Electronics 08 00292 v3 PDF

electronics
Review
A State-of-the-Art Survey on Deep Learning Theory
and Architectures
Md Zahangir Alom 1, *, Tarek M. Taha 1 , Chris Yakopcic 1 , Stefan Westberg 1 , Paheding Sidike 2 ,
Mst Shamima Nasrin 1 , Mahmudul Hasan 3 , Brian C. Van Essen 4 , Abdul A. S. Awwal 4 and
Vijayan K. Asari 1
1 Department of Electrical and Computer Engineering, University of Dayton, Dayton, OH 45469, USA;
ttaha1@udayton.edu (T.M.T.); cyakopcic1@udayton.edu (C.Y.); westbergs1@udayton.edu (S.W.);
nasrinm1@udayton.edu (M.S.N.); vasari1@udayton.edu (V.K.A.)
2 Department of Earth and Atmospheric Sciences, Saint Louis University, Saint Louis, MO 63108, USA;
sidike.paheding@slu.edu
3 Comcast Labs, Washington, DC 20005, USA; mahmud.ucr@gmail.com
4 Lawrence Livermore National Laboratory (LLNL), Livermore, CA 94550, USA;
vanessen1@llnl.gov (B.C.V.E.); awwal1@llnl.gov (A.A.S.A.)
* Correspondence: alomm1@udayton.edu

Received: 17 January 2019; Accepted: 31 January 2019; Published: 5 March 2019
Abstract: In recent years, deep learning has garnered tremendous success in a variety of application
domains. This new field of machine learning has been growing rapidly and has been applied to
most traditional application domains, as well as some new areas that present more opportunities.
Different methods have been proposed based on different categories of learning, including supervised,
semi-supervised, and un-supervised learning. Experimental results show state-of-the-art performance
using deep learning when compared to traditional machine learning approaches in the fields of
image processing, computer vision, speech recognition, machine translation, art, medical imaging,
medical information processing, robotics and control, bioinformatics, natural language processing,
cybersecurity, and many others. This survey presents a brief survey on the advances that have
occurred in the area of Deep Learning (DL), starting with the Deep Neural Network (DNN). The
survey goes on to cover Convolutional Neural Network (CNN), Recurrent Neural Network (RNN),
including Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU), Auto-Encoder (AE),
Deep Belief Network (DBN), Generative Adversarial Network (GAN), and Deep Reinforcement
Learning (DRL). Additionally, we have discussed recent developments, such as advanced variant
DL techniques based on these DL approaches. This work considers most of the papers published
after 2012 from when the history of deep learning began. Furthermore, DL approaches that have
been explored and evaluated in different application domains are also included in this survey. We
also included recently developed frameworks, SDKs, and benchmark datasets that are used for
implementing and evaluating deep learning approaches. There are some surveys that have been
published on DL using neural networks and a survey on Reinforcement Learning (RL). However,
those papers have not discussed individual advanced techniques for training large-scale deep learning
models and the recently developed method of generative models.
Keywords: deep learning; convolutional neural network (CNN); recurrent neural network (RNN);
auto-encoder (AE); restricted Boltzmann machine (RBM); deep belief network (DBN); generative
adversarial network (GAN); deep reinforcement learning (DRL); transfer learning
Electronics 2019, 8, 292; doi:10.3390/electronics8030292 www.mdpi.com/journal/electronics

Since the 1950s, a small subset of Artificial Intelligence (AI), often called Machine Learning (ML),
has revolutionized several fields in the last few decades. Neural Networks (NN) is a subfield of ML,
and it was this subfield that spawned Deep Learning (DL). Since its inception DL has been creating
ever larger disruptions, showing outstanding success in almost every application domain. Figure 1
shows the
Electronics 2019,taxonomy
8, 292 of AI. DL which uses either deep architectures of learning or hierarchical 2 of 66
learning approaches), is a class of ML developed largely from 2006 onward. Learning is a procedure
consisting of estimating the model parameters so that the learned model (algorithm) can perform a
1. Introduction
specific task. For example, in Artificial Neural Networks (ANN), the parameters are the weight
matrices.
SinceDL, on theaother
the 1950s, small hand,
subsetconsists of several
of Artificial layers(AI),
Intelligence in between
often calledthe Machine
input and output (ML),
Learning layer
Electronics 2019, 8, x FOR PEER REVIEW 3 of 67
which
has allows for several
revolutionized many fields
stagesinofthenon-linear information
last few decades. Neuralprocessing
Networks (NN) units iswith hierarchical
a subfield of ML,
architectures
and to be present that are exploited for feature learning and pattern classification [1,2].
1.1.2.itDeep
was this subfield that spawned
Semi-supervised Learning Deep Learning (DL). Since its inception DL has been creating
Learning methods based on representations of data can also be defined as
ever larger disruptions, showing outstanding success in almost every application domain. Figure 1 representation learning
shows Semi-supervised
[3]. Recent literature of
the taxonomy learning
states
AI. DL is
DLlearning
thatwhich based that deep
occurs
representation
uses either based on partially
learning
architectures involves
of learninglabeled datasets.
a hierarchy In some
of features
or hierarchical learningor
cases,
concepts, DRL and
where Generative
the Adversarial
high-level concepts Networks
can be (GAN)
defined are
from used
the as semi-supervised
low-level
approaches), is a class of ML developed largely from 2006 onward. Learning is a procedure consisting ones and learning
low-level
techniques.
concepts
of estimatingcanGANbe is discussed
thedefined
model in Section
from high-level
parameters so that7.the
Section
ones. In 8 surveys
some
learned DRL approaches.
articles,(algorithm)
model DL has been Additionally,
candescribed
perform aasspecificRNN,
a universal
task.
including
learning
For example,LSTM
approach and GRU,
in Artificial are
that is able used
to solve
Neural for semi-supervised
almost(ANN),
Networks all kinds learning
theofparameters as
problems in well.
aredifferent application
the weight matrices. domains.
DL, on
In other
the otherwords,
hand, DL is notof
consists task specific
several [4]. in between the input and output layer which allows for
layers
1.1.3. Deep Unsupervised Learning
many stages of non-linear information processing units with hierarchical architectures to be present
1.1. Type
that are of Deep Learning
exploited
Unsupervised for
learning Approaches
feature learning
systems areand
ones pattern
that canclassification
without the [1,2]. Learning
presence of datamethods
labels. Inbased on
this case,
representations
the agent learnsof data
the can
internal also be defined
representation asorrepresentation
important learning
features to [3]. Recent
discover
Deep learning approaches can be categorized as follows: Supervised, semi-supervised or literature
unknown states that
relationships
DL
or basedsupervised,
structure
partially representation
within the andlearning
input involves
data.
unsupervised. In aaddition,
Often hierarchy
clustering, ofdimensionality
there features or concepts,
is another where
reduction,
category of the generative
and
learning high-level
approach
concepts
techniques canare
be considered
defined
called Reinforcement from
Learningasthe low-level
unsupervised
(RL) or DeeponesRLand
learning low-level
(DRL) whichconcepts
approaches. can
There
are often be several
are defined
discussed from
members
under high-level
the of the
scope of
ones.
deep In some
learning articles,
family DL
that has
are been
good described
at clustering as a
anduniversal
non-linearlearning approach
dimensionality
semi-supervised or sometimes under unsupervised learning approaches. Figure 2 shows the pictorial that is
reduction,able to solve
including
almost
diagram. all kinds of
Auto-Encoders problems
(AE), in different
Restricted application
Boltzmann Machinesdomains.
(RBM), In and
otherthe
words, DL isdeveloped
recently not task specific
GAN.[4].In
addition, RNNs, such as LSTM and RL, are also used for unsupervised learning in many application
domains. Sections 6 and 7 discuss RNNs and LSTMs in detail.
1.1.4. Deep Reinforcement Learning (RL)

Deep Reinforcement Learning is a learning technique for use in unknown environments. DRL
began in 2013 with Google Deep Mind [5,6]. From then on, several advanced methods have been
proposed based on RL. Here is an example of RL: If environment samples inputs: 𝑥 ~𝜌 , agent
predict: 𝑦 = 𝑓(𝑥 ) , agent receive cost: 𝑐 ~𝑃(𝑐 |𝑥 , 𝑦 ) where P is an unknown probability
distribution, the environment asks an agent a question, and gives a noisy score as the answer.
Sometimes this approach is called semi-supervised learning as well. There are many semi-supervised
and un-supervised techniques that have been implemented based on this concept (in Section 8). In
RL, we do not have a straight forward loss function, thus making learning harder compared to
Figure supervised
Figure
traditional 1.1.The
Thetaxonomy
taxonomy of AI:
of AI.
approaches. AI. Artificial
AI: fundamental
The Artificial Intelligence;
Intelligence; ML:between
ML: Machine
differences Machine
Learning;
RLLearning;
NN:
andNeural NN: Neural
Networks;
supervised learning
Networks;
DL: Deep DL: Deep
Learning; Learning;
SNN: SpikingSNN: Spiking
Neural Neural
Networks. Networks.
are: First, you do not have full access to the function you are trying to optimize; you must query them
through
1.1. Type interaction,
of Deep Learning second, you are interacting with a state-based environment: Input 𝑥
andApproaches
1.1.1. Deep Supervised
depends on previous actions. Learning
Deep
Dependinglearning
Supervised approaches
upon
learning can be
theisproblem
a learning categorized
scope or space,
technique as one
that follows:
usescan Supervised,
decide
labeled Insemi-supervised
which
data. thetype
caseofofRL or partially
needs
supervised to DL
be
supervised, and unsupervised.
a task. If the In addition,
problem has athere
lot of is another
parameters category
to be of
approaches, the environment has a set of inputs and corresponding outputs (𝑥 , 𝑦 )~𝜌. For example,
applied for solving learning
optimized, DRL approach
is the bestcalled
way
Reinforcement
to go. input
if for If the xt, Learning
problem (RL)fewer
has
the intelligentor Deep RLpredicts
(DRL) for
parameters
agent which are often discussed
), the
𝑦 optimization,
= 𝑓(𝑥 under
a derivation
agent will the
freescope
receive RL of semi-supervised
value 𝑙(𝑦
approach
a loss , 𝑦 ).
is good.
or
An sometimes
Theexample
agent will under
of this unsupervised
thenisiteratively learning
annealing, modify approaches.
cross entropy methods,
the network Figure 2 shows
and SPSA.
parameters the pictorial diagram.
for a better approximation of the
desired outputs. After successful training, the agent will be able to get the correct answers to
questions from the environment. There are different supervised learning approaches for deep
leaning, including Deep Neural Networks (DNN), Convolutional Neural Networks (CNN),
Recurrent Neural Networks (RNN), including Long Short Term Memory (LSTM), and Gated
Recurrent Units (GRU). These networks will be described in details in the respective sections.
approaches.
Figure 2. Category of Deep Leaning approaches.
1.2. Feature Learning

A key difference between traditional ML and DL is in how features are extracted. Traditional
ML approaches use handcrafted engineering features by applying several feature extraction
algorithms, and then apply the learning algorithms. Additionally, other boosting approaches are
Electronics 2019, 8, 292 3 of 66
1.1.1. Deep Supervised Learning

Supervised learning is a learning technique that uses labeled data. In the case of supervised DL
approaches, the environment has a set of inputs and corresponding outputs ( xt , yt ) ∼ ρ. For example,
if for input xt, the intelligent agent predicts ŷt = f ( xt ), the agent will receive a loss value l (yt , ŷt ). The
agent will then iteratively modify the network parameters for a better approximation of the desired
outputs. After successful training, the agent will be able to get the correct answers to questions from
the environment. There are different supervised learning approaches for deep leaning, including Deep
Neural Networks (DNN), Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN),
including Long Short Term Memory (LSTM), and Gated Recurrent Units (GRU). These networks will
be described in details in the respective sections.
1.1.2. Deep Semi-supervised Learning

Semi-supervised learning is learning that occurs based on partially labeled datasets. In some cases,
DRL and Generative Adversarial Networks (GAN) are used as semi-supervised learning techniques.
GAN is discussed in Section 7. Section 8 surveys DRL approaches. Additionally, RNN, including
LSTM and GRU, are used for semi-supervised learning as well.
1.1.3. Deep Unsupervised Learning

Unsupervised learning systems are ones that can without the presence of data labels. In this case,
the agent learns the internal representation or important features to discover unknown relationships or
structure within the input data. Often clustering, dimensionality reduction, and generative techniques
are considered as unsupervised learning approaches. There are several members of the deep learning
family that are good at clustering and non-linear dimensionality reduction, including Auto-Encoders
(AE), Restricted Boltzmann Machines (RBM), and the recently developed GAN. In addition, RNNs,
such as LSTM and RL, are also used for unsupervised learning in many application domains. Sections 6
and 7 discuss RNNs and LSTMs in detail.
1.1.4. Deep Reinforcement Learning (RL)

Deep Reinforcement Learning is a learning technique for use in unknown environments. DRL
began in 2013 with Google Deep Mind [5,6]. From then on, several advanced methods have been
proposed based on RL. Here is an example of RL: If environment samples inputs: xt ∼ ρ, agent predict:
ŷt = f ( xt ), agent receive cost: ct ∼ P(ct | xt , ŷt ) where P is an unknown probability distribution, the
environment asks an agent a question, and gives a noisy score as the answer. Sometimes this approach
is called semi-supervised learning as well. There are many semi-supervised and un-supervised
techniques that have been implemented based on this concept (in Section 8). In RL, we do not have
a straight forward loss function, thus making learning harder compared to traditional supervised
approaches. The fundamental differences between RL and supervised learning are: First, you do not
have full access to the function you are trying to optimize; you must query them through interaction,
and second, you are interacting with a state-based environment: Input xt depends on previous actions.
Depending upon the problem scope or space, one can decide which type of RL needs to be applied
for solving a task. If the problem has a lot of parameters to be optimized, DRL is the best way to go.
If the problem has fewer parameters for optimization, a derivation free RL approach is good. An
example of this is annealing, cross entropy methods, and SPSA.
1.2. Feature Learning

A key difference between traditional ML and DL is in how features are extracted. Traditional ML
approaches use handcrafted engineering features by applying several feature extraction algorithms,
and then apply the learning algorithms. Additionally, other boosting approaches are often used where
several learning algorithms are applied to the features of a single task or dataset and a decision is
made according to the multiple outcomes from the different algorithms.
On the other hand, in the case of DL, the features are learned automatically and are represented
hierarchically in multiple levels. This is the strong point of DL against traditional machine learning
approaches. Table 1 shows the different feature-based learning approaches with different learning steps.
Table 1. Different feature learning approaches.
Approaches Learning Steps

Rule-based Input Hand-design features Output
Traditional Machine Learning Input Hand-design features Mapping from features Output
Representation Learning Input Features Mapping from features Output
Deep Learning Input Simple features Complex features Mapping from features Output
1.3. Why and When to apply DL

DL is employed in several situations where machine intelligence would be useful (see Figure 3):
• Absence of a human expert (navigation on Mars)

• Humans are unable to explain their expertise (speech recognition, vision, and language understanding)
• The solution to the problem changes over time (tracking, weather prediction, preference, stock,
price prediction)
• Solutions need to be adapted to the particular cases (biometrics, personalization).
• The problem size is too vast for our limited reasoning capabilities (calculation webpage ranks,
matching ads to Facebook, sentiment analysis).
The experimental results on ImageNet-2012

18 16.4
16
14
11.2
12
Erros (%)
10
7.4 6.7
8
6 5
3.57
4
2
0
2012 2013 2014 2014 2015
AlexNet[7] Clarifia[8] VGG-16 [9] GoogLeNet- ResNet- Human

19 [10] 152[11]
Model & Year
Figure 3. Accuracy
Accuracy for ImageNet classification challenge with different DL models.
At present, DL is being applied in almost all areas. As a result, this approach is often called a
Phone error rate (PER) in percentage(%)
universal learning approach.
1.4. The State-of-the-art Performance of RNN[23]

Segmental DL
There are some outstanding successes
Attention-based RNN[22]in the fields of computer vision and speech recognition as
discussed below:
RNN
(a). Image classification ontransducer[19]
ImageNet dataset. One of the large-scale problems is named Large
Scale Visual Recognition Challenge (LSVRC). CNN and its variants as one of the DL branches showed
Ensemble DNN/CNN/RNN[21]
DCNN [20]
CTC[19]
CDNN w. Hater.s Pooling[18]

3.57
4
2
0
2012 2013 2014 2014 2015
Electronics 2019, 8, 292 AlexNet[7] Clarifia[8] VGG-16 [9] GoogLeNet- ResNet- Human 5 of 66
19 [10] 152[11]
Model & Year
state-of-the-art accuracy on the ImageNet task [7–12]. The following graph shows the success story of
DL techniques overtime on ImageNet-2012 challenge. Figure 3 shows that ResNet-152 has achieved
Figure 3. Accuracy for ImageNet classification challenge with different DL models.
3.57% error rate which outperformed human accuracy.
Phone error rate (PER) in percentage(%)
Segmental RNN[23]
Attention-based RNN[22]
RNN transducer[19]
Ensemble DNN/CNN/RNN[21]
DCNN [20]
CTC[19]
CDNN w. Hater.s Pooling[18]
DSC with 2nd pass[16]
End-to-end DL [17]
Discriminative segmetal cascade (DSC)[16]
Deep Segmental NN[15]
Boundary-factored SCRF[14]
First-pass SCRF [13]
0 5 10 15 20 25 30 35
Figure 4. Phone error rate (PER) for TIMIT Acoustic-Phonetic Continuous Speech Corpus dataset [13–23].
Figure 4. Phone error rate (PER) for TIMIT Acoustic-Phonetic Continuous Speech Corpus dataset [13–
23]. Automatic speech recognition. The initial success in the field of speech recognition on the
(b).
popular TIMIT dataset (common data set are generally used for evaluation) was with small-scale
(b). Automatic speech recognition. The initial success in the field of speech recognition on the
recognition tasks [24]. The TIMIT Acoustic-Phonetic continuous speech Corpus contains 630 speakers
popular TIMIT dataset (common data set are generally used for evaluation) was with small-scale
from eight major dialects of American English, where each speaker reads 10 sentences. Figure 4
recognition tasks [24]. The TIMIT Acoustic-Phonetic continuous speech Corpus contains 630 speakers
summarizes
Electronics 2019,the
8, x error rates,
FOR PEER including these early results and is measured as a percent phone error6 rate
REVIEW of 67
from eight major dialects of American English, where each speaker reads 10 sentences. Figure 4
(PER) over the last 20 years. The bar graph clearly shows that the recently developed DL approaches
summarizes(top
approaches the error
therates, including these early results and is measured as a percent phone error
(top of the graph) of perform graph)
better perform
compared better
to anycompared to any
other previous other previous
machine machine learning
learning approaches on the
rate (PER)
approaches over the last 20
on the TIMIT dataset. years. The bar graph clearly shows that the recently developed DL
TIMIT dataset.
Some example applications are are shown
shown in
in Figures
Figures 55 and
and 6.
6.
Road 1 Road 2
Road 3
Figure 5. Data-driven
Figure 5. Data-driven traffic
traffic forecasting. Using the
forecasting. Using the dynamics
dynamics of
of the
the traffic
traffic flow
flow (roads
(roads 1,
1, 2,
2, and
and 3)
3) to
to
capture the spatial dependency using by Diffusion Convolutional Recurrent Neural Network
capture the spatial dependency using by Diffusion Convolutional Recurrent Neural Network [25]. [25].
Semantic segmentation [27]

Lymphocyte detection [26]
Road 3
Figure 5. Data-driven traffic forecasting. Using the dynamics of the traffic flow (roads 1, 2, and 3) to
capture the spatial dependency using by Diffusion Convolutional Recurrent Neural Network [25].
Semantic segmentation [27]

Lymphocyte detection [26]
Nuclei segmentation [28] Blood cell classification [29]
Example images
Figure6.6. Example
Figure imageswhere DL DL
where is applied successfully
is applied and achieved
successfully and state-of-the-art performance.
achieved state-of-the-art
The images were taken from the corresponding references [26–29].
performance. The images were taken from the correspond ding references.
1.5. Why DL?

1.5. Why DL?
1.5.1. Universal Learning Approach
1.5.1. Universal Learning Approach
The DL approach is sometimes called universal learning because it can be applied to almost any
The DL approach
application domain. is sometimes called universal learning because it can be applied to almost any
application domain.
1.5.2. Robust
1.5.2. Robust
Deep learning approaches do not require the precisely designed feature. Instead, optimal features
Deep learning learned
are automatically approaches dotask
for the notatrequire
hand. Asthea result,
precisely designed feature.
the robustness Instead,
to natural optimal
variations of the
features are isautomatically
input data achieved. learned for the task at hand. As a result, the robustness to natural
variations of the input data is achieved.
1.5.3. Generalization
1.5.3. Generalization
The same DL approach can be used in different applications or with different data types. This
The same
approach DL approach
is often can be
called transfer used inIndifferent
learning. addition,applications
this approachor iswith different
helpful wheredata types. This
the problem does
approach
not haveissufficient
often called transfer
available learning.
data. In addition,
There are a numberthis approachthat
of literatures is helpful where thethis
have discussed problem
concept
(See Section 4).
1.5.4. Scalability
The DL approach is highly scalable. Microsoft invented a deep network known as ResNet [11].
This network contains 1202 layers and is often implemented at a supercomputing scale. There is a big
initiative at Lawrence Livermore National Laboratory (LLNL) in developing frameworks for networks
like this, which can implement thousands of nodes [24].
1.6. Challenges of DL
There are several challenges for DL:
The DL approach is highly scalable. Microsoft invented a deep network known as ResNet [11].
This network contains 1202 layers and is often implemented at a supercomputing scale. There is a big
initiative at Lawrence Livermore National Laboratory (LLNL) in developing frameworks for
networks like this, which can implement thousands of nodes [24].
1.6. Challenges of DL
• Big data
There areanalytics using DL for DL:
several challenges
• • Big data analytics
Scalability of DL approaches using DL
• • toScalability
Ability of DLwhich
generate data approaches
is important where data is not available for learning the system
• Ability to generate data
(especially for computer vision task, whichsuchis as
important where data is not available for learning the
inverse graphics).
• system (especially
Energy efficient techniquesfor forcomputer vision task,
special purpose suchincluding
devices, as inversemobile
graphics).
intelligence, FPGAs,
•
and so on. Energy efficient techniques for special purpose devices, including mobile intelligence,
• FPGAs, and so on.
Multi-task and transfer learning or multi-module learning. This means learning from different
• Multi-task
domains and transfer
or with different modelslearning
together.or multi-module learning. This means learning from
• Dealing with causality in learning. different models together.
different domains or with
• Dealing with causality in learning.
Most
Most ofof the
the above-mentioned
above-mentioned challenges have already been considered by the DL community.
Firstly,
Firstly,for
forthe
thebig
bigdata
dataanalytics
analyticschallenge,
challenge,there
thereisisaagood
goodsurvey
surveythat
thatwas
wasconducted
conductedin in2014
2014[30].
[30]. In
In
this paper, the authors explained details on how DL can deal with different criteria,
this paper, the authors explained details on how DL can deal with different criteria, including including volume,
velocity,
volume, variety,
velocity, and veracity
variety, andofveracity
the big data problem.
of the big dataThe authorsThe
problem. alsoauthors
showedalsodifferent
showedadvantages
different
of DL approaches
advantages of DL when dealing with
approaches whenbig dealing
data problems [31,32].
with big dataFigure 7 clearly
problems demonstrates
[31,32]. Figure 7that the
clearly
performance of traditional ML approaches shows better performance for lesser amounts
demonstrates that the performance of traditional ML approaches shows better performance for lesser of input data.
As the amount
amounts of inputof data
data.increases beyondofadata
As the amount certain number,
increases the performance
beyond of traditional
a certain number, machine
the performance
learning approaches
of traditional machinebecomes steady,
learning whereasbecomes
approaches DL approaches
steady,increase
whereas with
DLrespect to theincrease
approaches increment
withof
the amount
respect to theof increment
data. of the amount of data.
Figure 7. The performance of deep learning with respect to the amount of data.
Figure 7. The performance of deep learning with respect to the amount of data.
Secondly, in most of the cases for solving large-scale problems, the solution is being implemented
Secondly, in most of the cases for solving large-scale problems, the solution is being
on High-Performance Computing (HPC) system (super-computing, cluster, sometimes considered
implemented on High-Performance Computing (HPC) system (super-computing, cluster, sometimes
cloud computing) which offers immense potential for data-intensive business computing. As data
considered cloud computing) which offers immense potential for data-intensive business computing.
explodes in velocity, variety, veracity, and volume, it is getting increasingly difficult to scale compute
As data explodes in velocity, variety, veracity, and volume, it is getting increasingly difficult to scale
performance using enterprise-class servers and storage in step with the increase. Most of the articles
compute performance using enterprise-class servers and storage in step with the increase. Most of
considered all the demands and suggested efficient HPC with heterogeneous computing systems. In
the articles considered all the demands and suggested efficient HPC with heterogeneous computing
one example, Lawrence Livermore National Laboratory (LLNL) has developed a framework which
systems. In one example, Lawrence Livermore National Laboratory (LLNL) has developed a
is called Livermore Big Artificial Neural Networks (LBANN) for large-scale implementation (in
framework which is called Livermore Big Artificial Neural Networks (LBANN) for large-scale
super-computing scale) for DL which clearly supplants the issue of scalability of DL [24].
Thirdly, generative models are another challenge for deep learning. One example is the GAN,
which is an outstanding approach for data generation for any task which can generate data with the
same distribution [33]. Fourthly, multi-task and transfer learning which we have discussed in Section 7.
Fourthly, there is a lot of research that has been conducted on energy efficient deep learning approaches
with respect to network architectures and hardwires. Section 10 discusses this issue.
Can we make any uniform model that can solve multiple tasks in different application domains?
As far as the multi-model system is concerned, one article from Google titled One Model To Learn Them
All [34] is a good example. This approach can learn from different application domains, including
ImageNet, multiple translation tasks, Image captioning (MS-COCO dataset), speech recognition
corpus and English parsing task. We will be discussing most of the challenges and respective solutions
through this survey. There are some other multi-task techniques that have been proposed in the last
few years [35–37].
Finally, a learning system with causality has been presented, which is a graphical model that
defines how one may infer a causal model from data. Recently a DL based approach has been proposed
for solving this type of problem [38]. However, there are other many challenging issues have been
solved in the last few years which were not possible to solve efficiently before this revolution. For
example, image or video captioning [39], style transferring from one domain to another domain using
GAN [40], text to image synthesis [41], and many more [42].
There are some surveys that have been conducted recently in the DL field [43–46]. These
papers survey on DL and its revolution, but they did not address the recently developed generative
model called GAN [33]. In addition, they discuss little RL and did not cover recent trends of DRL
approaches [1,44]. In most of the cases, the surveys that have been conducted are on different
DL approaches individually. There is a good survey which is based on Reinforcement Learning
approaches [46,47]. Another survey exists on transfer learning [48]. One survey has been conducted
on neural network hardware [49]. However, the main objective of this work is to provide an overall
idea on deep learning and its related fields, including deep supervised (e.g., DNN, CNN, and RNN),
unsupervised (e.g., AE, RBM, GAN) (sometimes GAN also used for semi-supervised learning tasks)
and DRL. In some cases, DRL is considered to be a semi-supervised or an unsupervised approach. In
addition, we have considered the recently developing trends in this field and applications which are
developed based on these techniques. Furthermore, we have included the framework and benchmark
datasets which are often used for evaluating deep learning techniques. Moreover, the name of the
conferences and journals are also included which are considered by this community for publishing
their research articles.
The rest of the paper has been organized in the following ways: The detailed surveys of DNNs are
discussed in Section 2, Section 3 discusses on CNN. Section 4 describes different advanced techniques
for efficient training of DL approaches. Section 5 discusses RNNs. AEs and RBMs are discussed in
Section 6. GANs with applications are discussed in Section 7. RL is presented in Section 8. Section 9
explains transfer learning. Section 10 presents energy efficient approaches and hardwires for DL.
Section 11 discusses deep learning frameworks and standard development kits (SDK). The benchmarks
for different application domains with web links are given in Section 12. The conclusions are made in
Section 13.
2. Deep Neural Network
2.1. The History of DNN

A brief history of neural networks highlighting key events is, as shown in Figure 8. Computational
neurobiology has conducted significant research on constructing computational models of artificial
neurons. Artificial neurons, which try to mimic the behavior of the human brain, are the fundamental
component for building ANNs. The basic computational element (neuron) is called a node (or unit)
which receives inputs from external sources and has some internal parameters (including weights and
biases that are learned during training) which produce outputs. This unit is called a perceptron. The
fundamental of ANN is discussed in References [1,3].
(including
Electronics weights
2019, and biases that are learned during training) which produce outputs. This unit
8, 292 is
9 of 66
called a perceptron. The fundamental of ANN is discussed in References [1,3].
1943 - present
•1943: McCulloch & Pitts show that neurons can be combined to construct a Turing
machine [50].
•1958: Rosenblatt shows that perceptron’s will converge if what they are trying to
learn can be represented [51].
•1969: Minsky & Papert show the limitations of perceptron’s, killing research in
neural networks for a decade [52].
•1985: The backpropagation algorithm by Geoffrey Hinton et al. [53] revitalizes the
field.
•1988: Neocognitron: a hierarchical neural network capable of visual pattern
recognition [54].
•1998: CNNs with Backpropagation for document analysis by Yan LeCun [55].
•2006: The Hinton lab solves the training problem for DNNs [56,57].
•2012 - pressent: A variety of deep learning algorithms are increasingly emerging.
Figure 8. The history of deep learning development [50–57].

Figure 8. The history of deep learning development [50–57].
ANNs or general NNs consist of Multilayer Perceptron’s (MLP) which contain one or more
ANNs
hidden or general
layers NNs consist
with multiple hiddenofunits
Multilayer Perceptron’s
(neurons) in them. (MLP) which
For details oncontain one orsee
MLP, please more
in
hidden layers with
References [1,3,53] multiple hidden units (neurons) in them. For details on MLP, please see in
References [1,3,53]
2.2. Gradient Descent
2.2. Gradient Descent
The gradient descent approach is a first-order optimization algorithm which is used for finding
The gradient descent approach is a first-order optimization algorithm which is used for finding
the local minima of an objective function. This has been used for training ANNs in the last couple of
the local minima of an objective function. This has been used for training ANNs in the last couple of
decades successfully [1,53].
decades successfully [1,53].
2.3. Stochastic Gradient Descent (SGD)
2.3. Stochastic Gradient Descent (SGD)
Since a long training time is the main drawback for the traditional gradient descent approach,
Since a long training time is the main drawback for the traditional gradient descent approach, the
the SGD approach is used for training Deep Neural Networks (DNN) [1,58].
SGD approach is used for training Deep Neural Networks (DNN) [1,58].
2.4. Back-Propagation
2.4. (BP)
Back-Propagation (BP)
DNN is
DNN is trained
trained with
with the
the popular
popular Back-Propagation
Back-Propagation (BP)
(BP) algorithm
algorithm with
with SGD
SGD [47,53].
[47,53]. In
In the
the case
case
of MLPs,
of MLPs, we
we can
can easily
easily represent
represent NN
NN models
models using
using computation
computation graphs
graphs which
which are
are directive
directive acyclic
acyclic
graphs. For that representation of DL, we can use the chain-rule to efficiently calculate the gradient
graphs. For that representation of DL, we can use the chain-rule to efficiently calculate the gradient
from the
from the top
top to
to the
the bottom
bottom layers
layers with
with BP,
BP,as
asshown
shownininReferences
References[53,59–63].
[53,59–63].
2.5. Momentum
Momentum is a method which helps to accelerate the training process with with the
the SGD
SGD approach.
approach.
The main
The main idea behind
behind it is to use the moving average of the gradient instead of using only the current
the gradient.
real value of the gradient. We
We can
can express
express this
this with
with the
the following
followingequation
equationmathematically:
mathematically:
vt𝑣==γγv𝑣t−1 −−η𝜂∇ F (θt−1),),

∇ ℱ(𝜃 (1)
(1)
𝜃 =𝜃 + 𝑣, (2)
θ t = θ t −1 + v t , (2)
here γ is the momentum and 𝜂 is the learning rate for the tth round of training. Other popular
here γ is the momentum and η is the learning rate for the tth round of training. Other popular
approaches have been introduced during the last few years which are explained in section IX under
approaches have been introduced during the last few years which are explained in Section 9 under
approaches. The main advantage of using momentum during training is to
the scope of optimization approaches.
prevent
prevent the network from gettingstuck
the network from getting stuckininlocal
localminimum.
minimum.The Thevalues
valuesofof
momentum γ∈
areare
momentum ∈ (0,1].
γ (0,1]. It is
It
noted that a higher momentum value overshoots its minimum, possibly making the network unstable.
is noted that a higher momentum value overshoots its minimum, possibly making the network
In general, γ is set to 0.5 until the initial learning stabilizes and is then increased to 0.9 or higher [60].
Electronics 2019, 8, 292 10 of 66
2.6. Learning Rate (η )

The learning rate is an important component for training DNN. The learning rate is the step size
considered during training which makes the training process faster. However, selecting the value of
the learning rate is sensitive. For example: If you choose a larger value for η, the network may start
diverging instead of converging. On the other hand, if you choose a smaller value for η, it will take
more time for the network to converge. In addition, it may easily get stuck in local minima. The typical
solution to this problem is to reduce the learning rate during training [64].
There are three common approaches used for reducing the learning rate during training: Constant,
factored, and exponential decay. First, we can define a constant ζ which is applied to reduce the
learning rate manually with a defined step function. Second, the learning rate can be adjusted during
training with the following equation:
t
η t = η0 β e , (3)
where ηt is the tth round learning rate, η0 is the initial learning rate, and β is the decay factor with a
value between the range of (0, 1).
The step function format for exponential decay is:
t
η t = η0 β b e c . (4)
The common practice is to use a learning rate decay of β = 0.1 to reduce the learning rate by a
factor of 10 at each stage.
2.7. Weight Decay

Weight decay is used for training deep learning models as an L2 regularization approach, which
helps to prevent overfitting the network and model generalization. L2 regularization for F (θ, x ) can
be defined as,
Ω = k θ k2 , (5)
1
ε̂(F (θ, x ), y) = ε(F (θ, x ), y) + λ Ω. (6)
2
The gradient for the weight θ is:
∂ 21 λΩ
= λ·θ. (7)
∂θ
General practice is to use the value λ = 0.0004. A smaller λ will accelerate training.
Other necessary components for efficient training, including data preprocessing and augmentation,
network initialization approaches, batch normalization, activation functions, regularization with dropout,
and different optimization approaches (as discussed in Section 4).
In the last few decades, many efficient approaches have been proposed for better training of
deep neural networks. Before 2006, attempts taken at training deep architectures failed: Training a
deep supervised feed-forward neural network tended to yield worse results (both in training and
in test error) than shallow ones (with 1 or 2 hidden layers). Hinton’s revolutionary work on DBNs
spearheaded a change in this in 2006 [56,59].
Due to their composition, many layers of DNNs are more capable of representing highly varying
nonlinear functions compared to shallow learning approaches [62–65]. Moreover, DNNs are more
efficient for learning because of the combination of feature extraction and classification layers. The
following sections discuss in detail about different DL approaches with necessary components.
Electronics 2019, 8, 292 11 of 66
3. Convolutional Neural Network (CNN)
3.1.Electronics
CNN Overview
2019, 8, x FOR PEER REVIEW 11 of 67
This network structure was first proposed by Fukushima in 1988 [54]. It was not widely
3.1. CNN Overview
used, however, due to limits of computation hardware for training the network. In the 1990s,
LeCun et Thisal. network
[55] appliedstructure was first proposed
a gradient-based learningby algorithm
FukushimatoinCNNs 1988 [54].
andItobtained
was not successful
widely used, results
however, due to limits of computation hardware for training the network.
for the handwritten digit classification problem. After that, researchers further improved CNNs In the 1990s, LeCun et al.
and[55] appliedstate-of-the-art
reported a gradient-based learning
results algorithm
in many to CNNs
recognition and obtained
tasks. CNNs have successful
severalresults for the over
advantages
handwritten digit classification problem. After that, researchers further improved CNNs and
DNNs, including being more like the human visual processing system, being highly optimized in the
reported state-of-the-art results in many recognition tasks. CNNs have several advantages over
structure for processing 2D and 3D images, and being effective at learning and extracting abstractions
DNNs, including being more like the human visual processing system, being highly optimized in the
of 2D features.
structure The max pooling
for processing 2D and 3D layer of CNNs
images, is effective
and being effective inatabsorbing
learning and shape variations.
extracting Moreover,
abstractions
composed of sparse
of 2D features. Theconnections
max poolingwith layertied weights,
of CNNs CNNs in
is effective have significantly
absorbing fewer parameters
shape variations. Moreover, than a
fully connected network of similar size. Most of all, CNNs are trained with
composed of sparse connections with tied weights, CNNs have significantly fewer parameters than the gradient-based learning
algorithm and suffernetwork
a fully connected less from the diminishing
of similar size. Most of gradient
all, CNNs problem. Given
are trained withthat
thethe gradient-based
gradient-based
algorithm
learningtrains the whole
algorithm network
and suffer to minimize
less from an errorgradient
the diminishing criterionproblem.
directly,Given
CNNs canthe
that produce
gradient-highly
based algorithm
optimized weights.trains the whole network to minimize an error criterion directly, CNNs can produce
highly
Figure optimized
9 showsweights.
the overall architecture of CNNs consists of two main parts: Feature extractors
and a classifier.9 shows
Figure In the the overall
feature architecture
extraction of CNNs
layers, each consists
layer ofofthe two main parts:
network Feature
receives theextractors
output from
its immediate previous layer as its input and passes its output as the input to the nextfrom
and a classifier. In the feature extraction layers, each layer of the network receives the output its The
layer.
immediate previous layer as its input and passes its output as the input to the next layer. The CNN
CNN architecture consists of a combination of three types of layers: Convolution, max-pooling, and
architecture consists of a combination of three types of layers: Convolution, max-pooling, and
classification. There are two types of layers in the low and middle-level of the network: Convolutional
classification. There are two types of layers in the low and middle-level of the network: Convolutional
layers and max-pooling layers. The even numbered layers are for convolutions and the odd-numbered
layers and max-pooling layers. The even numbered layers are for convolutions and the odd-
layers are for max-pooling
numbered layers are for operations.
max-pooling The output The
operations. nodes of thenodes
output convolution and max-pooling
of the convolution and max- layers
are pooling
groupedlayersinto aare2Dgrouped
plane called
into a 2D plane called feature mapping. Each plane of a layer is usually the
feature mapping. Each plane of a layer is usually derived from
combination
derived from of one
the or more planes
combination of previous
of one layers.ofThe
or more planes nodes layers.
previous of a plane
The are connected
nodes of a planeto are
a small
region of eachtoconnected
connected planesof
a small region of each
the previous
connected layer. Each
planes of node of the convolution
the previous layer. Eachlayer
nodeextracts
of the the
convolution
features from the layer extracts
input the features
images from the operations
by convolution input images onbythe convolution
input nodes.operations on the input
nodes.
Figure 9. The
Figure 9. Theoverall
overallarchitecture
architecture of
of the
the Convolutional Neural
Convolutional Neural Network
Network (CNN)
(CNN) includes
includes an input
an input
layer, multiple
layer, multiple alternating
alternatingconvolution
convolution and max-poolinglayers,
and max-pooling layers,
oneone fully-connected
fully-connected layerlayer
and and
one one
classification layer.
classification layer.
Higher-level
Higher-levelfeatures
featuresarearederived
derived from
from features propagatedfrom
features propagated from lower
lower level
level layers.
layers. As theAs the
features
features propagatetotothe
propagate thehighest
highest layer
layerororlevel,
level,thethe
dimensions
dimensions of features are reduced
of features depending
are reduced on
depending
the size
on the size of
ofthethekernel
kernelforfor
thethe
convolutional
convolutional and andmax-pooling
max-pooling operations respectively.
operations However,
respectively. the
However,
the number of feature maps usually increased for representing better features of the input for
number of feature maps usually increased for representing better features of the input images images
for ensuring
ensuringclassification accuracy.
classification TheThe
accuracy. output of theof
output last
thelayer
lastoflayer
the CNN is used
of the CNNasisthe input
used astothe
a fully
input to
connected network which is called classification layer. Feed-forward neural networks have been used
a fully connected network which is called classification layer. Feed-forward neural networks have
as the classification layer as they have better performance [56,64]. In the classification layer, the
extracted features are taken as inputs with respect to the dimension of the weight matrix of the final
Electronics 2019, 8, 292 12 of 66
been used as the classification layer as they have better performance [56,64]. In the classification
layer, the extracted features are taken as inputs with respect to the dimension of the weight matrix of
the final neural network. However, the fully connected layers are expensive in terms of network or
learning parameters. Nowadays, there are several new techniques, including average pooling and
global average pooling that is used as an alternative of fully-connected networks. The score of the
respective class is calculated in the top classification layer using a soft-max layer. Based on the highest
score, the classifier gives output for the corresponding classes. Mathematical details on different layers
of CNNs are discussed in the following section.
3.1.1. Convolutional Layer

In this layer, feature maps from previous layers are convolved with learnable kernels. The output
of the kernels goes through a linear or non-linear activation function, such as sigmoid, hyperbolic
tangent, Softmax, rectified linear, and identity functions) to form the output feature maps. Each of the
output feature maps can be combined with more than one input feature map. In general, we have that
 
x lj = f  ∑ xil −1 ∗ klij + blj , (8)

ieM j
where x lj is the output of the current layer, xil −1 is the previous layer output, klij is the kernel for the
present layer, and blj are the biases for the current layer. M j represents a selection of input maps. For
each output map, an additive bias b is given. However, the input maps will be convolved with distinct
kernels to generate the corresponding output maps. The output maps finally go through a linear
or non-linear activation function (such as sigmoid, hyperbolic tangent, Softmax, rectified linear, or
identity functions).
3.1.2. Sub-sampling Layer

The subsampling layer performs the down sampled operation on the input maps. This is
commonly known as the pooling layer. In this layer, the number of input and output feature maps
does not change. For example, if there are N input maps, then there will be exactly N output maps.
Due to the down sampling operation, the size of each dimension of the output maps will be reduced,
depending on the size of the down sampling mask. For example, if a 2 × 2 down sampling kernel is
used, then each output dimension will be half of the corresponding input dimension for all the images.
This operation can be formulated as
xlj = down xlj −1 , (9)
where down(.) represents a sub-sampling function. Two types of operations are mostly performed in
this layer: Average pooling or max-pooling. In the case of the average pooling approach, the function
usually sums up over N × N patches of the feature maps from the previous layer and selects the
average value. On the other hand, in the case of max-pooling, the highest value is selected from the
N × N patches of the feature maps. Therefore, the output map dimensions are reduced by n times. In
some special cases, each output map is multiplied with a scalar. Some alternative sub-sampling layers
have been proposed, such as fractional max-pooling layer and sub-sampling with convolution. These
are explained in Section 4.6.
3.1.3. Classification Layer

This is the fully connected layer which computes the score of each class from the extracted
features from a convolutional layer in the preceding steps. The final layer feature maps are represented
as vectors with scalar values which are passed to the fully connected layers. The fully connected
feed-forward neural layers are used as a soft-max classification layer. There are no strict rules on
Electronics 2019, 8, 292 13 of 66
the number of layers which are incorporated in the network model. However, in most cases, two to
Electronics 2019, 8, x FOR PEER REVIEW
four layers have been observed in different architectures, including LeNet [55], AlexNet [7], and13VGG
of 67
Net [9]. As the fully connected layers are expensive in terms of computation, alternative approaches
layer and the average pooling layer which help to reduce the number of parameters in the network
have been proposed during the last few years. These include the global average pooling layer and the
significantly.
average pooling layer which help to reduce the number of parameters in the network significantly.
In the backward propagation through the CNNs, the fully connected layer updates following
In the backward propagation through the CNNs, the fully connected layer updates following
the general approach of fully connected neural networks (FCNN). The filters of the convolutional
the general approach of fully connected neural networks (FCNN). The filters of the convolutional
layers are updated by performing the full convolutional operation on the feature maps between the
layers are updated by performing the full convolutional operation on the feature maps between the
convolutional layer and its immediate previous layer. Figure 10 shows the basic operations in the
convolutional layer and its immediate previous layer. Figure 10 shows the basic operations in the
convolution and sub-sampling of an input image.
convolution and sub-sampling of an input image.
Feature maps after performing convolution and pooling operations.

Figure 10. Feature
3.1.4. Network Parameters and Required Memory for CNN

3.1.4. Network Parameters and Required Memory for CNN
The number of computational parameters is an important metric to measure the complexity of a
The number of computational parameters is an important metric to measure the complexity of
deep learning model. The size of the output feature maps can be formulated as follows:
a deep learning model. The size of the output feature maps can be formulated as follows:
((𝑁
N − 𝐹)
F)
𝑀=
M = +
+1,
1, (10)
(10)
𝑆
S
where N 𝑁 refers
refersto
to the
the dimensions
dimensions of the input feature maps, 𝐹
feature maps, refersto
F refers to the
the dimensions
dimensions of the filters
or the receptive field, 𝑀 refers
receptive field, M refers to the dimensions of output feature maps, and 𝑆S stands
output feature maps, and stands for the stride
length. Padding is typically
typically applied during the convolution
convolution operations to ensure the input and output
feature map have the same dimensions.
dimensions. The
The amount
amount ofof padding
padding depends
depends on on the
the size
size of
of the
the kernel.
kernel.
Equation 17 is used for determining the number of rows and columns for padding.
(17) is used for determining the number of rows and columns for padding.
𝑃 = (𝐹 − 1)/2, (11)
P = ( F − 1)/2, (11)
here 𝑃 is the amount of padding and 𝐹 refers to the dimension of the kernels. Several criteria are
here P is the
considered foramount of padding
comparing and FHowever,
the models. refers to in
themost
dimension
of the of the kernels.
cases, Several
the number criteria
of network
are considered
parameters andfor
thecomparing
total amounttheofmodels.
memoryHowever, in most
are considered. of number
The the cases,
of the number (𝑃𝑎𝑟𝑚
parameters ) of
of network
𝑙parameters andcalculated
layer is the the total amount of the
based on memory equation: The number of parameters ( Parml ) of l th
are considered.
following
layer is the calculated based on the following equation:
𝑃𝑎𝑟𝑚 = (𝐹 × 𝐹 × 𝐹𝑀 ) × 𝐹𝑀 . (12)
If bias is added with the weights, l = (F
Parmthen × above
the F × FM l −1 ) × FM
equation canl . be written as follows: (12)
𝑃𝑎𝑟𝑚 =
If bias is added with the weights, (𝐹 the
then × (𝐹above
+ 1) × 𝐹𝑀 )can
equation × 𝐹𝑀
be ,written as follows: (13)
here the total number of parameters of 𝑙 lathe yer can be represented with 𝑃 , 𝐹𝑀 is for the total
number of output feature maps, Parmand ( F × (isF the
l =𝐹𝑀 × FM
+ 1)total l −1 ) × of
number FM l,
input (13)
feature maps or channels.
For example, let’s assume the 𝑙 layerth has 𝐹𝑀 = 32 input features maps, 𝐹𝑀 = 64 output
here the total number of parameters of l lathe yer can be represented with Pl , FMl is for the total
feature maps, and the filter size is 𝐹 = 5. In this case, the total number of parameters with a bias for
number of output feature maps, and FMl −1 is the total number of input feature maps or channels.
this layer: 𝑃𝑎𝑟𝑚 = (5 × 5 × 33) × 64 = 528,000. Thus, the amount of memory (𝑀𝑒𝑚 ) needs for the
For example, let’s assume the l th layer has FMl −1 = 32 input features maps, FMl = 64 output feature
operations of the 𝑙 layer can be expressed as
𝑀𝑒𝑚 = (𝑁 × 𝑁 × 𝐹𝑀 ). (14)
3.2. Popular CNN Architectures

Electronics 2019, 8, 292 14 of 66
maps, and the filter size is F = 5. In this case, the total number of parameters with a bias for this layer:
Parml = (5 × 5 × 33) × 64 = 528, 000. Thus, the amount of memory (Meml ) needs for the operations
of the l th layer can be expressed as
Electronics 2019, 8, x FOR PEER REVIEW

Meml = ( Nl × Nl × FMl ). (14)
14 of 67
3.2. Popular CNN Architectures

In this section, several popular state-of-the-art CNN architectures will be examined. In general,
mostIndeepthis section, severalneural
convolutional popular state-of-the-art
networks are made CNN ofarchitectures
a key set ofwill be examined.
basic In general,
layers, including the
most deep convolutional
convolution neural networks
layer, the sub-sampling are made
layer, dense of and
layers, a keytheset of basiclayer.
soft-max layers,Theincluding
architecturesthe
convolution
typically consist layer,
of the sub-sampling
stacks layer, dense layers,
of several convolutional andmax-pooling
layers and the soft-maxlayerslayer.followed
The architectures
by a fully
typically
connectedconsist of stacks
and SoftMax of several
layers at theconvolutional
end. Some examples layers and max-pooling
of such models are layers
LeNet followed by a fully
[55], AlexNet [7],
connected
VGG Net [9], andNiN
SoftMax layers
[66] and allatconvolutional
the end. Some(All examples
Conv) of such
[67]. models
Other are LeNet
alternatives [55],
and AlexNet
more [7],
efficient
VGG Net [9], NiN [66] and all convolutional (All Conv) [67]. Other
advanced architectures have been proposed, including DenseNet [68], FractalNet [69], GoogLeNet alternatives and more efficient
advanced
with Inceptionarchitectures
units have been proposed,
[10,70,71], and Residual including DenseNet
Networks [68],
[11]. FractalNet
The [69], GoogLeNet
basic building componentswith
Inception
(convolution unitsand[10,70,71],
pooling)and Residual
are almost theNetworks
same across [11]. Thearchitectures.
these basic buildingHowever,
components some (convolution
topological
and pooling) are
differences are almost
observedthe same
in the across these architectures.
modern deep learning However, some topological
architectures. Of the many differencesDCNNare
observed
architectures,in the modern[7],
AlexNet deep
VGG learning architectures.
[9], GoogLeNet Of the Dense
[10,70,71], many DCNN
CNN [68] architectures,
and FractalNet AlexNet [7],
[69] are
VGG [9], GoogLeNet
generally considered the [10,70,71], Densearchitectures
most popular CNN [68] and FractalNet
because [69]
of their are generallyperformance
state-of-the-art considered the on
most popular
different architectures
benchmarks because
for object of their state-of-the-art
recognition tasks. Amongperformance
all of these on different some
structures, benchmarks
of the
for object recognition
architectures tasks.especially
are designed Among all for of these structures,
large-scale some(such
data analysis of theasarchitectures
GoogLeNet are anddesigned
ResNet),
especially
whereas the forVGG
large-scale
network datais analysis
considered (such as GoogLeNet
a general and ResNet),
architecture. Some ofwhereas the VGG network
the architectures are dense is
considered
in terms of aconnectivity,
general architecture. Some of the
such as DenseNet architectures
[68]. are dense
Fractal Network is aninalternative
terms of connectivity,
of ResNet model.such as
DenseNet [68]. Fractal Network is an alternative of ResNet model.
3.2.1. LeNet (1998)
3.2.1. LeNet (1998)
Although LeNet was proposed in the 1990s, limited computation capability and memory
capacity made LeNet
Although was proposed
the algorithm difficult in to
theimplement
1990s, limiteduntilcomputation
about 2010 [55].capability
LeCun and memory
et al. capacity
[55], however,
made the algorithm difficult to implement until about 2010 [55].
proposed CNNs with the back-propagation algorithm and experimented on handwritten digit LeCun et al. [55], however, proposed
CNNs
dataset with the back-propagation
to achieve state-of-the-art accuracy.algorithm Theand experimented
proposed on handwritten
CNN architecture digit dataset
is well-known as LeNet-to
achieve state-of-the-art accuracy. The proposed CNN architecture is
5 [55]. The basic configuration of LeNet-5 is as follows (see Figure 11): Two convolutions (conv) well-known as LeNet-5 [55].
The basic
layers, two configuration
sub-sampling of layers,
LeNet-5 twois as follows
fully (see Figure
connected layers,11):
andTwoan convolutions
output layer with (conv)the layers, two
Gaussian
sub-sampling
connection. The layers,
total two fullyof
number connected
weights and layers, and anand
Multiply output layer with(MACs)
Accumulates the Gaussian
are 431 connection.
k and 2.3
The total number
M, respectively. of weights and Multiply and Accumulates (MACs) are 431 k and 2.3 M, respectively.
As
As computational
computational hardware
hardware started
started improving
improving in in capability,
capability, CNNs
CNNs stated
stated becoming
becoming popular
popular asas
an effective learning approach in the computer vision and machine
an effective learning approach in the computer vision and machine learning communities. learning communities.
Figure 11. The architecture of LeNet.

Figure 11. The architecture of LeNet.
3.2.2. AlexNet (2012)
3.2.2. AlexNet (2012)
In 2012, Alex Krizhevesky and others proposed a deeper and wider CNN model compared to
LeNetInand
2012,
wonAlex
the Krizhevesky
most difficultand others challenge
ImageNet proposedfor
a deeper and wider
visual object CNN model
recognition compared
called the ImageNetto
LeNet and won the most difficult ImageNet challenge for visual object recognition
Large Scale Visual Recognition Challenge (ILSVRC) in 2012 [7]. AlexNet achieved state-of-the-art called the
ImageNet Large Scale Visual Recognition Challenge (ILSVRC) in 2012 [7]. AlexNet achieved state-of-
the-art recognition accuracy against all the traditional machine learning and computer vision
approaches. It was a significant breakthrough in the field of machine learning and computer vision
for visual recognition and classification tasks and is the point in history where interest in deep
learning increased rapidly.
Electronics 2019, 8, 292 15 of 66
recognition accuracy against all the traditional machine learning and computer vision approaches.
It was a significant breakthrough in the field of machine learning and computer vision for visual
recognition and classification tasks and is the point in history where interest in deep learning
increased rapidly.
The architecture of AlexNet is shown in Figure 12. The first convolutional layer performs
convolution and max-pooling with Local Response Normalization (LRN) where 96 different receptive
filters are used that are 11 × 11 in size. The max pooling operations are performed with 3 × 3 filters
with a stride
3 filters with size of 2.size
a stride Theofsame
2. Theoperations are performed
same operations in the second
are performed layer with
in the second layer5× 5 filters.
with 5×5
× 3 filters
3filters. are used in the third, fourth, and fifth convolutional layers with 384, 384, and
3 × 3 filters are used in the third, fourth, and fifth convolutional layers with 384, 384, and 296 feature
296
maps
feature respectively. Two fullyTwo
maps respectively. connected (FC) layers(FC)
fully connected are used with
layers aredropout
used withfollowed by afollowed
dropout Softmax layer
by a
at the end.
Softmax Two
layer at networks
the end. Twowithnetworks
similar structure andstructure
with similar the sameandnumber of feature
the same number maps are trained
of feature maps
in parallel for this model. Two new concepts, Local Response Normalization
are trained in parallel for this model. Two new concepts, Local Response Normalization (LRN) (LRN) and dropout,
and
are introduced
dropout, in this network.
are introduced LRN can be
in this network. LRNapplied
can beinapplied
two different
in twoways: Firstways:
different applying
First on single
applying
channel
on singleorchannel
feature maps, where
or feature maps, × N patch
an Nwhere is selected
an N×N from
patch is the same
selected fromfeature mapfeature
the same and normalized
map and
based
normalized based on the neighborhood values. Second, LRN can be applied across thefeature
on the neighborhood values. Second, LRN can be applied across the channels or maps
channels or
(neighborhood along the third dimension but a single pixel or location).
feature maps (neighborhood along the third dimension but a single pixel or location).
Figure 12. The architecture of AlexNet: Convolution, max-pooling, Local Response Normalization
Figure 12. The architecture of AlexNet: Convolution, max-pooling, Local Response Normalization
(LRN) and fully connected (FC) layer.
(LRN) and fully connected (FC) layer.
AlexNet has three convolution layers and two fully connected layers. When processing the
AlexNet has three convolution layers and two fully connected layers. When processing the
ImageNet dataset, the total number of parameters for AlexNet can be calculated as follows for the
ImageNet dataset, the total number of parameters for AlexNet can be calculated as follows for the
first layer: Input samples are 224 × 224 × 3, filters (kernels or masks) or a receptive field that has
first layer: Input samples are 224×224×3, filters (kernels or masks) or a receptive field that has a size
a size 11, the stride is 4, and the output of the first convolution layer is 55 × 55 × 96. According
11, the stride is 4, and the output of the first convolution layer is 55×55×96. According to the equations
to the equations in Section 3.1.4, we can calculate that this first layer has 290,400 (55 × 55 × 96)
in section 3.1.4, we can calculate that this first layer has 290400 (55×55×96) neurons and 364 (11 ×11×3
neurons and 364 (11 ×11 × 3 = 363 + 1 bias) weights. The parameters for the first convolution layer
= 363 + 1 bias) weights. The parameters for the first convolution layer are 290400×364 = 105,705,600.
are 290,400 × 364 = 105,705,600. Table 2 shows the number of parameters for each layer in millions.
Table 2 shows the number of parameters for each layer in millions. The total number of weights and
The total number of weights and MACs for the whole network are 61M and 724M, respectively.
MACs for the whole network are 61M and 724M, respectively.
3.2.3. ZFNet / Clarifai (2013)
3.2.3. ZFNet / Clarifai (2013)
In 2013, Matthew Zeiler and Rob Fergue won the 2013 ILSVRC with a CNN architecture which
was anIn extension
2013, Matthew Zeiler and
of AlexNet. The Rob Fergue
network waswon theZFNet
called 2013 ILSVRC
[8], afterwith
the aauthors’
CNN architecture which
names. As CNNs
was an extension of AlexNet. The network was called ZFNet [8], after the authors’
are expensive computationally, an optimum use of parameters is needed from a model complexitynames. As CNNs
are expensive
point of view. computationally, an optimum
The ZFNet architecture is an use of parameters
improvement is neededdesigned
of AlexNet, from a model complexity
by tweaking the
point of parameters
network view. The ZFNet architecture
of the latter. ZFNet is an7improvement
uses of AlexNet,
× 7 kernels instead of 11 ×designed bytotweaking
11 kernels the
significantly
network parameters of the latter. ZFNet uses 7x7 kernels instead of 11x11 kernels
reduce the number of weights. This reduces the number of network parameters dramatically and to significantly
reduce theoverall
improves number of weights.
recognition This reduces the number of network parameters dramatically and
accuracy.
improves overall recognition accuracy.
3.2.4. Network in Network (NiN)

This model is slightly different from the previous models where a couple of new concepts are
introduced [66]. The first concept is to use multilayer perception convolution, where convolutions
are performed with 1×1 filter that help to add more nonlinearity in the models. This helps to increase
the depth of the network, which can then be regularized with dropout. This concept is used often in
the bottleneck layer of a deep learning model.
Electronics 2019, 8, 292 16 of 66
3.2.4. Network in Network (NiN)

This model is slightly different from the previous models where a couple of new concepts are
introduced [66]. The first concept is to use multilayer perception convolution, where convolutions are
performed with 1 × 1 filter that help to add more nonlinearity in the models. This helps to increase the
depth of the network, which can then be regularized with dropout. This concept is used often in the
bottleneck layer of a deep learning model.
The second concept is to use Global Average Pooling (GAP) as an alternative of fully connected
layers. This helps to reduce the number of network parameters significantly. GAP changes the
network structure significantly. By applying GAP on a large feature map, we can generate a final low
dimensional
Electronics 2019, feature vector
8, x FOR PEER without reducing the dimension of the feature maps.
REVIEW 16 of 67
3.2.5. VGGNET
3.2.5. VGGNET (2014)
(2014)
The Visual
The Visual Geometry
Geometry GroupGroup (VGG),
(VGG), was was the
the runner-up
runner-up of of the
the 2014
2014 ILSVRC
ILSVRC [9].[9]. The
The main
main
contribution of
contribution ofthis
thiswork
workis is that
that it shows
it shows thatthat the depth
the depth of a network
of a network is a critical
is a critical componentcomponent to
to achieve
achieve better recognition or classification accuracy in CNNs. The VGG
better recognition or classification accuracy in CNNs. The VGG architecture consists of two architecture consists of two
convolutional layers
convolutional layersboth
both of of
which
whichuse use
the ReLU activation
the ReLU function.
activation Following
function. the activation
Following function
the activation
is a single max pooling layer and several fully connected layers also using a ReLU
function is a single max pooling layer and several fully connected layers also using a ReLU activation activation function.
The final layer
function. of thelayer
The final modelofisthe a Softmax
model layeris a for classification.
Softmax layer forInclassification.
VGG-E [9] theInconvolution
VGG-E [9]filter
the
size is changed
convolution tosize
filter × changed
a 3 is 3 filter with
to a a3 stride of with
× 3 filter 2. Three VGG-E
a stride of 2. [9] models,
Three VGG-E VGG-11, VGG-16,
[9] models, and
VGG-11,
VGG-19;
VGG-16, were proposedwere
and VGG-19; the models
proposed hadthe11, models
16, and 19
hadlayers respectively.
11, 16, and 19 layersThe respectively.
VGG networkThe model
VGG is
shown inmodel
network Figureis13.
shown in Figure 13.
Figure
Figure 13. The
13. The basic
basic building
building block
block of VGG
of VGG network:
network: Convolution
Convolution (Conv)
(Conv) andand FC fully
FC for for fully connected
connected layers.
layers.
All versions of the VGG-E models ended the same with three fully connected layers. However,
All
the numberversions of the VGG-E
of convolution models
layers ended
varied the same
VGG-11 with three
contained fully connected
8 convolution layers.
layers, However,
VGG-16 had
theconvolution
13 number of convolution layers varied
layers, and VGG-19 VGG-11
had 16 contained
convolution 8 convolution
layers. VGG-19, the layers,
mostVGG-16 had 13
computational
convolution layers,
expensive model, and VGG-19
contained had 16and
138Mweights convolution
had 15.5 Mlayers.
MACs.VGG-19, the most computational
expensive model, contained 138Mweights and had 15.5 M MACs.
3.2.6. GoogLeNet (2014)
3.2.6. GoogLeNet (2014)
GoogLeNet, the winner of ILSVRC 2014 [10], was a model proposed by Christian Szegedy of
Google with the objective
GoogLeNet, the winner of of
reducing
ILSVRCcomputation complexity
2014 [10], was comparedby
a model proposed to Christian
the traditional CNN.
Szegedy of
The proposed
Google method
with the was
objective ofto incorporate
reducing Inceptioncomplexity
computation Layers that had variable
compared to thereceptive fields,
traditional CNN.which
The
were created
proposed by different
method kernel sizes.Inception
was to incorporate These receptive
Layers fields created
that had operations
variable receptive that captured
fields, whichsparse
were
correlation
created by patterns
differentinkernel
the new feature
sizes. mapreceptive
These stack. fields created operations that captured sparse
The initial
correlation concept
patterns in theofnew
thefeature
Inception
maplayer
stack.can be seen in Figure 14. GoogLeNet improved
state-of-the-art recognition accuracy using a stack of Inception layers, seen in Figure 15. The difference
between the naïve inception layer and final Inception Layer was the addition of 1 × 1 convolution
kernels. These kernels allowed for dimensionality reduction before computationally expensive layers.
GoogLeNet consisted of 22 layers in total, which was far greater than any network before it. Later
improved version of this network is proposed in [71]. However, the number of network parameters
Figure 14. Inception layer: Naive version.

convolution layers, and VGG-19 had 16 convolution layers. VGG-19, the most computational
expensive model, contained 138Mweights and had 15.5 M MACs.
3.2.6. GoogLeNet (2014)

Electronics 2019, 8, 292 17 of 66
GoogLeNet, the winner of ILSVRC 2014 [10], was a model proposed by Christian Szegedy of
Google with the objective of reducing computation complexity compared to the traditional CNN. The
proposed method
GoogLeNet wasmuch
used was to incorporate
lower thanInception LayersAlexNet
its predecessor that hadorvariable receptive fields,
VGG. GoogLeNet had 7Mwhich were
network
created by when
parameters different kernelhad
AlexNet sizes.
60MThese receptive
and VGG-19 fields
138M. Thecreated operations
computations that captured
for GoogLeNet alsosparse
were
correlation patterns in the new feature map stack.
1.53G MACs far lower than that of AlexNet or VGG.
7M network parameters when AlexNet had 60M and VGG-19 138M. The computations for
GoogLeNet also were 1.53G MACs far14.
Figure
Figure
lower
14. than layer:
Inception
Inception
that ofNaive
layer: AlexNet or VGG.
Naive version.
version.
The initial concept of the Inception layer can be seen in Figure 14. GoogLeNet improved state-
of-the-art recognition accuracy using a stack of Inception layers, seen in Figure 15. The difference
between the naïve inception layer and final Inception Layer was the addition of 1x1 convolution
kernels. These kernels allowed for dimensionality reduction before computationally expensive
layers. GoogLeNet
7M network consisted of 22
parameters layers
when in total,
AlexNet which
had 60M and was far 138M.
VGG-19 greater
Thethan any network
computations for before it.
GoogLeNet also were 1.53G MACs far lower than that of AlexNet or VGG.
Later improved version of this network is proposed in [71]. However, the number of network
parameters GoogLeNet used was much lower than its predecessor AlexNet or VGG. GoogLeNet had
Figure 15. Inception layer with dimension reduction.
3.2.7. Residual Network (ResNet in 2015)

3.2.7. Residual Network (ResNet in 2015)
The winner of ILSVRC 2015 was the Residual Network architecture, ResNet [11]. Resnet was
The winner of ILSVRC 2015 was the Residual Network architecture, ResNet [11]. Resnet was
developed by Kaiming He with Figure the intent of designing
15. Inception ultra-deep
layer with dimension reduction. networks that did not suffer from
developed by Kaiming He with the intent of designing ultra-deep networks that did not suffer from
the vanishing gradient problem that predecessors had. ResNet is developed with many different
the vanishing gradient
3.2.7. problem
Residual Network thatinpredecessors
(ResNet 2015) had. ResNet is developed with many different
numbers of layers; 34, 50,101, 152, and even 1202. The popular ResNet50 contained 49 convolution
numbers of layers; The 34,
winner50,101, 152,2015
of ILSVRC andwas even 1202. The
the Residual popular
Network ResNet50
architecture, ResNetcontained
[11]. Resnet was 49 convolution
layers and 1 fully connected layer at the end ofdesigning
the network. The total that
number of weights and MACs
layers and 1 developed
fully connected
by Kaiming layer at the
He with end of
the intent of the network. The
ultra-deep total number
networks did notof weights
suffer from and MACs
for the wholethenetwork
vanishingare 25.5M
gradient and 3.9M
problem respectively.
that predecessors had. ResNet is developed with many different
for the whole network are 25.5M and 3.9M respectively.
numbers
The basic blockofdiagram
layers; 34, of
50,101,
the 152, and even
ResNet 1202. The popular
architecture is shownResNet50 contained
in Figure 16.49ResNet
convolution is a traditional
The basic block
layers and diagram of the
1 fully connected ResNet
layer architecture
at the end of the network.is The
shown in Figure
total number 16. ResNet
of weights and MACs is a traditional
feedforward network with a residual connection. The output of a residual layer can be defined based
feedforward for
network
the wholewith
th a residual connection. The output of a residual layer can be defined based
network are 25.5M and 3.9M respectively.
on the outputs ofThe( −
l (𝑙 1)block
basic which
diagramcomes
of the from
ResNetthe previous
architecture layerindefined
is shown Figure 16.as xl −1 is. aFtraditional
ResNet ( x ) is the output
on the outputs of − 1) which comes from the previous layer defined as 𝑥 l.−1ℱ( 𝑥 ) is the
feedforward network with a residual connection. The output of a
after performing various operations (e.g., convolution with different size of filters, Batchresidual layer can be defined based
Normalization
output after on
performing
the outputs ofvarious operations
(𝑙 − 1) which comes from (e.g.,
the convolution with as
previous layer defined different
𝑥 . ℱ( 𝑥size ) isof
thefilters, Batch
(BN) followed by an activation function, such as a ReLU on x l − 1 ). The final output of
size 𝑥of filters,
residualthe unit
Normalization (BN)
output followed
after performing byvarious
an activation
operations function, such with
(e.g., convolution as adifferent
ReLU on ). The final output of
Batch
is xl which can be defined
Normalization (BN)with the by
followed following
an activationequation:
function, such as a ReLU on 𝑥 ). The final output of
residualthe unit is 𝑥 which can be defined with the following equation:
residualthe unit is 𝑥 which can be defined with the following equation:
x𝑥l == ℱ(
𝑥 F
= (ℱ(x𝑥𝑥l −1)))++
+𝑥 𝑥xl.−1.. (15) (15)
(15)
16. Basic
FigureFigure diagram
16. Basic ofthe
diagram of the Residual
Residual block.block.
The residual network consists of several basic residual blocks. However, the operations in the
residual block can be varied depending on the different architecture of residual networks [11]. The
wider version of the residual network was proposed by Zagoruvko el at. [72], another improved
Figure
residual network approach 16. Basic
known diagram
as aggregated of thetransformation
residual Residual block.[73]. Recently, some other
variants of residual models have been introduced based on the Residual Network architecture [74–
76]. Furthermore, there are several advanced architectures that are combined with Inception and
Electronics 2019, 8, 292 18 of 66
The residual network consists of several basic residual blocks. However, the operations in the
residual block can be varied depending on the different architecture of residual networks [11]. The
wider version of the residual network was proposed by Zagoruvko et al. [72], another improved
residual network approach known as aggregated residual transformation [73]. Recently, some other
variants of residual models have been introduced based on the Residual Network architecture [74–76].
Furthermore, there are several advanced architectures that are combined with Inception and Residual
units. The
Electronics basic
2019, 8, x conceptual diagram of Inception-Residual unit is shown in the following Figure
FOR PEER REVIEW 18 17.
of 67
Figure 17.
Figure The basic
17. The basic block
block diagram
diagram for
for Inception
Inception Residual
Residual unit.
unit.
Figure 17. The basic block diagram for Inception Residual unit.
Mathematically, this concept can be represented as
Mathematically, this concept can be represented as
Mathematically, this concept can be represented × K
as
×
F ℱ(xl3𝑥−×1×3 ⨀ 𝑥xl5×−×15) ++𝑥 xl −, 1 ,

xl 𝑥= = (16)
(16)
𝑥 = ℱ( 𝑥 ⨀𝑥 )+ 𝑥 , (16)
where the symbol J ⨀ refers the concentration operations between two outputs from the 3×3 and 5×5
where the
where symbol ⨀ refers
the symbol refersthetheconcentration
concentrationoperations operationsbetween
betweentwo twooutputs
outputsfrom from the 3× 3 and
filters. After that, the convolution operation is performed with 1×1 filters. Finally, the
the 3×3 and
outputs 5×5
are
5 × 5 filters.
filters. After
Afterthe
added with that, that, the convolution
the convolution
inputs of this block operation
operation
of 𝑥 . The
is
is conceptperformed
performed with 1
with 1×1 block
of Inception
× 1 filters.
filters.with Finally,
Finally, the outputs
the connections
residual outputs are
are added
added with
with
is introduced theinthe inputs
inputs of this
of this
the Inception-v4
block
block of of𝑥 xl −. 1The
architecture
. The concept
concept
[71]. The of
ofInception
Inceptionblock
improved
block withresidual
version with
residualconnections
connections
of the Inception-Residual
is introduced
is introduced in in the Inception-v4
theproposed
Inception-v4 architecture
architecture [71]. [71]. The
The improved
improved version
version of of the
the Inception-Residual
Inception-Residual
network were also [76,77].
network were also proposed
network were also proposed [76,77]. [76,77].
3.2.8. Densely Connected Network (DenseNet)
3.2.8. Densely
3.2.8. Densely Connected
Connected Network
Network (DenseNet)
(DenseNet)
DenseNet developed by Gao et al. in 2017 [68], which consists of densely connected CNN layers,
DenseNet developed
DenseNet developed by by Gao
Gao etet al.
al. in
in 2017
2017 [68], which which consists
consists of ofdensely
denselyconnected
connected CNNCNN layers,
layers,
the outputs of each layer are connected with all[68], successor layers in a dense block [68]. Therefore, it is
the
the outputs
outputs of each layer are connected with all successor layers in a dense block [68]. Therefore, it is
is
formed withofdense
each connectivity
layer are connected
between with
the all successor
layers layersit in
rewarding thea name
dense DenseNet.
block [68]. This
Therefore,
conceptit is
formed with dense
formed dense connectivitybetween betweenthe thelayers
layersrewarding
rewardingititthe thename
nameDenseNet.
DenseNet. This concept
efficientwith
for featureconnectivity
reuse, which dramatically reduces network parameters. DenseNet This concept
consists is
of
is efficient
efficient for
for feature
feature reuse,
reuse, which
which dramatically
dramatically reduces
reduces network
network parameters.
parameters. DenseNet
DenseNet consists
consists of
of
several dense blocks and transition blocks, which are placed between two adjacent dense blocks. The
several dense
several dense blocks
blocks andand transition
transition blocks,
blocks, which
which are are placed
placed between
between twotwo adjacent
adjacent dense
dense blocks.
blocks. The
The
conceptual diagram of a dense block is shown in Figure 18.
conceptual diagram
conceptual diagram of of aa dense
dense block
block is is shown
shown in in Figure
Figure 18.
18.
Figure 18. A 4-layer Dense block with a growth rate of 𝑘 = 3.

Figure 18.
Figure A 4-layer
18. A 4-layer Dense
Dense block
block with growth rate of 𝑘k = 3.
with aa growth
Each layer takes all the preceding feature maps as input. When deconstructing Figure 19, the 𝑙
Each layerall
layer received takes
the all the preceding
feature maps from feature maps
previous as input.
layers of 𝑥 When as input:Figure 19, the 𝑙
, 𝑥 , 𝑥 deconstructing
⋯𝑥
layer received all the feature maps from previous layers of 𝑥 , 𝑥 , 𝑥 ⋯ 𝑥 as input:
𝑥 = 𝐻 ( 𝑥 ,𝑥 ,𝑥 ⋯𝑥 ), (17)
𝑥 = 𝐻 ( 𝑥 ,𝑥 ,𝑥 ⋯𝑥 ), (17)
where 𝑥 , 𝑥 , 𝑥 ⋯ 𝑥 are the concatenated features for layers 0, ⋯ ⋯ , 𝑙 − 1 and 𝐻 (∙) is considered
where 𝑥 , 𝑥 , 𝑥 ⋯ 𝑥 are
as a single tensor. It performs thethree
concatenated
differentfeatures for layers
consecutive operations: , 𝑙 − 1 and 𝐻 (∙) is considered
0, ⋯ ⋯Batch-Normalization (BN) [78],
as a single tensor. It performs three different consecutive operations: Batch-Normalization
followed by a ReLU [70] and a 3 × 3 convolution operation. In the transaction block, (BN)1[78],
×1
Electronics 2019, 8, 292 19 of 66
Each layer takes all the preceding feature maps as input. When deconstructing Figure 19, the l th
layer received all the feature maps from previous layers of x0 , x1 , x2 · · · xl −1 as input:
xl = Hl ([ x0 , x1 , x2 · · · xl −1 ]), (17)
where [ x0 , x1 , x2 · · · xl −1 ] are the concatenated features for layers 0, · · · · · · , l − 1 and Hl (·) is
Thisasarchitecture
considered a single tensor. is anItadvanced
performsand threealternative architectureoperations:
different consecutive of ResNet model, which is efficient
Batch-Normalization
for[78],
(BN) designing
followed large
by amodels
ReLU [70]withand
nominal
a 3 × 3depth, but shorter
convolution paths In
operation. forthe
thetransaction
propagation of gradient
block, 1×1
during training [69]. This concept is based on drop-path which is another
convolutional operations are performed with BN followed by a 2 × 2 average pooling layer. This new regularization approach
for making
model large networks.accuracy
shows state-of-the-art As a result,
with thisa concept helps
reasonable to enforce
number speed versus
of network accuracy
parameters fortradeoffs.
object
The basic block
recognitions tasks. diagram of FractalNet is shown in Figure 19.
Figure TheThe
19. 19.
Figure detailed FractalNet
detailed module
FractalNet on the
module leftleft
on the andand
FractalNet on the
FractalNet right.
on the right.
3.2.9. FractalNet (2016)
3.3. CapsuleNet
This architecture is an advanced and alternative architecture of ResNet model, which is efficient
CNNs large
for designing are an effective
models withmethodology
nominal depth,for but
detecting
shorterfeatures
paths forof the
an propagation
object and achieving good
of gradient
recognition performance compared to state-of-the-art handcrafted feature detectors. There
during training [69]. This concept is based on drop-path which is another regularization approach for are limits
to CNNs,
making large which are As
networks. thata result,
it doesthis
notconcept
take into account
helps special
to enforce relationships,
speed perspective,
versus accuracy size,
tradeoffs. Theand
orientation, of features. For example, if you have
basic block diagram of FractalNet is shown in Figure 19. a face image, it does not matter the placement of
different components (nose, eye, mouth, etc.) of the faces neurons of a CNN will wrongly active and
3.3.recognition
CapsuleNet as a face without considering special relationships (orientation, size). Now, imagine a
neuron which contains the likelihood with properties of features (perspective, orientation, size etc.).
CNNs are an effective methodology for detecting features of an object and achieving good
This special type of neurons, capsules, can detect face efficiently with distinct information. The
recognition performance compared to state-of-the-art handcrafted feature detectors. There are limits
capsule network consists of several layers of capsule nodes. The first version of capsule network
to CNNs, which are that it does not take into account special relationships, perspective, size, and
(CapsNet) consisted of three layers of capsule nodes in an encoding unit.
orientation, of features. For example, if you have a face image, it does not matter the placement of
This architecture for MNIST (28×28) images, the 256 9×9 kernels are applied with a stride 1, so
different components (nose, eye, mouth, etc.) of the faces neurons of a CNN will wrongly active and
the output is (28 − 9 + 1 = 20) with 256 feature maps. Then the outputs are fed to the primary
recognition as a face without considering special relationships (orientation, size). Now, imagine a
capsule layer which is a modified convolution layer that generates an 8-D vector instead of a scalar.
neuron which contains the likelihood with properties of features (perspective, orientation, size etc.).
In the first convolutional layer, 9×9 kernels are applied with stride 2, the output dimension is
This special type of neurons, capsules, can detect face efficiently with distinct information. The capsule
((20 − 9)/2 + 1 = 6). The primary capsules are used 8×32 kernels which generates 32×8×6×6 (32
network consists of several layers of capsule nodes. The first version of capsule network (CapsNet)
groups for 8 neurons with 6×6 size).
consisted of three layers of capsule nodes in an encoding unit.
This architecture for MNIST (28 × 28) images, the 256 9 × 9 kernels are applied with a stride 1,
so the output is (28 − 9 + 1 = 20) with 256 feature maps. Then the outputs are fed to the primary
capsule layer which is a modified convolution layer that generates an 8-D vector instead of a scalar.
In the first convolutional layer, 9 × 9 kernels are applied with stride 2, the output dimension is
Electronics 2019, 8, 292 20 of 66

((20 − 9)/2 + 1 = 6). The primary capsules are used 8 × 32 kernels which generates 32 × 8 × 6 × 6
(32 groups for 8 neurons with 6 × 6 size).
The entire encoding and decoding processes of CapsNet is shown in Figures 20 and 21, respectively.
We used a max-pooling layer in CNN often that can handle translation variance. Even if a feature
moves if it is still under a max pooling window it can be detected. As the capsule contains the weighted
sum of features from the previous layer, therefore this approach is capable of detecting overlapped
Electronics which
features 2019, 8, xisFOR PEER REVIEW
important for segmentation and detection tasks. 20 of 67
Figure 20. A CapsNet encoding unit with 3 layers. The instance of each class is represented with a
vector of a capsule in DigitCaps layer that is used for calculating classification loss. The weights
between the primary capsule layer and DigitCaps layer are represented with 𝑊 .
The entire encoding and decoding processes of CapsNet is shown in Figures 20 and 21,
respectively. We used a max-pooling layer in CNN often that can handle translation variance. Even
if a feature moves if it is still under a max pooling window it can be detected. As the capsule contains
Figure
Figure 20.
the weighted 20.sum CapsNet
A CapsNet encoding
encoding
of features unit
fromunit with 33 layers.
the with
previous layers. The
The
layer, instanceof
instance
therefore ofeach
this each classisisrepresented
class
approach isrepresented with
with
capable of a
detecting
avector
vectorofofa acapsule
capsuleininDigitCaps
DigitCapslayer
layer that
that is
is used
used for
for calculating
calculating classification
classification loss.
loss. The
The weights
weights
overlapped features which is important for segmentation and detection tasks.
between
between the
the primary
primary capsule
capsule layer
layer and
and DigitCaps
DigitCaps layer
layer are
arerepresented withW𝑊
representedwith ij . .
The entire encoding and decoding processes of CapsNet is shown in Figures 20 and 21,
respectively. We used a max-pooling layer in CNN often that can handle translation variance. Even
if a feature moves if it is still under a max pooling window it can be detected. As the capsule contains
the weighted sum of features from the previous layer, therefore this approach is capable of detecting
overlapped features which is important for segmentation and detection tasks.
Figure 21. The decoding unit where a digit is reconstructed from DigitCaps layer representation. The
Figure 21. The decoding unit where a digit is reconstructed from DigitCaps layer representation. The
Euclidean distance is used minimizing the error between the input sample and the reconstructed
Euclidean distance is used minimizing the error between the input sample and the reconstructed
sample from the sigmoid layer. True labels are used for reconstruction target during training.
sample from the sigmoid layer. True labels are used for reconstruction target during training.
In the traditional CNN, a single cost function is used to evaluate the overall error which propagates
In the
backward traditional
during training.CNN, a single
However, costcase,
in this function is usedbetween
if the weight to evaluate the overall
two neurons error
is zero, which
then the
activation of a neuron is not propagated from that neuron. The signal is routed with respect to theis
propagates backward during training. However, in this case, if the weight between two neurons
zero, then
feature the activation
parameters ratherof a neuron
than is not
a one size fitspropagated from that
all cost function neuron.dynamic
in iterative The signal is routed
routing withwith
the
Figure
respect to 21. The
the decoding
feature unit where
parameters a digit
rather is reconstructed
than a one size from
fits DigitCaps
all cost layer representation.
function in iterative The
dynamic
agreement. For details about this architecture, please see Reference [79]. This new CNN architecture
Euclidean
routing with distance
the is used For
agreement. minimizing about
the error
thisbetween the input sample and the reconstructed
provides state-of-the-art accuracydetails
for handwritten architecture,
digit recognition pleaseonsee Reference
MNIST. [79]. This
However, fromnew
an
sample from
CNN architecture the sigmoid layer. True labels are used for reconstruction target during training.
application point ofprovides
view, this state-of-the-art
architecture is accuracy for handwritten
more suitable digit recognition
for segmentation and detectionon MNIST.
tasks
However,
compare tofrom an application
classification tasks. point of view, this architecture is more suitable for segmentation and
In the traditional CNN, a single cost function is used to evaluate the overall error which
detection tasks compare to classification tasks.
propagates
3.4. Comparison backward during
of Different training. However, in this case, if the weight between two neurons is
Models
zero, then the activation of
3.4. Comparison of Different Models a neuron is not propagated from that neuron. The signal is routed with
Theto
respect comparison
the featureofparameters
recently proposed
rather models
than a based
one sizeon error,
fits allnetwork parameters,
cost function and a maximum
in iterative dynamic
number of connections
comparison are
of given
recently in Table
proposed2.
routing with the agreement. For details about this architecture, please see Reference [79]. Thisand
The models based on error, network parameters, newa
maximum number of connections are given in Table 2.
CNN architecture provides state-of-the-art accuracy for handwritten digit recognition on MNIST.
However, from an application point of view, this architecture is more suitable for segmentation and
Table 2. The top-5% errors with computational parameters and macs for different deep CNN models.
detection tasks compare to classification tasks.
3.4. Comparison of Different Models

Electronics 2019, 8, 292 21 of 66
Table 2. The top-5% errors with computational parameters and macs for different deep CNN models.
Methods LeNet-5 [54] AlexNet [7] OverFeat (fast) [8] VGG-16 [9] GoogLeNet [10] ResNet-50(v1) [11]
Top-5 errors n/a 16.4 14.2 7.4 6.7 5.3
Input size 28 × 28 227 × 227 231 × 231 224 × 224 224 × 224 224 × 224
Number of Conv Layers 2 5 5 16 21 50
Filter Size 5 3,5,11 3,7 3 1,3,5,7 1,3,7
Number of Feature Maps 1,6 3–256 3–1024 3–512 3–1024 3–1024
Stride 1 1,4 1,4 1 1,2 1,2
Number of Weights 26 k 2.3 M 16 M 14.7 M 6.0 M 23.5 M
Number of MACs 1.9 M 666 M 2.67 G 15.3 G 1.43 G 3.86 G
Number of FC layers 2 3 3 3 1 1
Number of Weights 406 k 58.6 M 130 M 124 M 1M 1M
Number of MACs 405 k 58.6 M 130 M 124 M 1M 1M
Total Weights 431 k 61 M 146 M 138 M 7M 25.5 M
Total MACs 2.3 M 724 M 2.8 G 15.5 G 1.43 G 3.9 G
3.5. Other DNN Models

There are many other network architectures, such as fast region-based CNN [80] and Xception [81],
which are popular in the computer vision community. In 2015 a new model was proposed using
recurrent convolution layers named Recurrent Convolution Neural Network or RCNN [82]. The
improved version of this network is a combination of the two most popular architectures in the
Inception network and Recurrent Convolutional Network, Inception Convolutional Recurrent Neural
Networks (IRCNN) [83]. IRCNN provided better accuracy compared RCNN and inception network
with almost identical network parameters. Visual Phase Guided CNN (ViP CNN) is proposed with
phase guided message passing a structure (PMPS) to build connections between relational components,
which show better speed up and recognition accuracy [84]. Look up based CNN [85] is a fast,
compact, and accurate model enabling efficient inference. In 2016 the architecture known as a fully
convolutional network (FCN) was proposed for segmentation tasks where it is now commonly used
in [27]. Other recently proposed CNN models include pixel net [86], a deep network with stochastic
depth, deeply-supervised networks, and ladder network [87–89]. Additional, CNN architecture models
are explained in [90]. Some articles are published on do deep nets really need to be deep [91–93].
There are some articles published on fitNet hits [94], initialization method [95], deep versus wide
net [96], training on DL on large training set [97], graph processing [98], energy efficient network
architectures [99,100].
3.6. Applications of CNNs
3.6.1. CNNs for Solving A Graph Problem

Learning graph data structures is a common problem with various applications in data mining
and machine learning tasks. DL techniques have made a bridge in between the machine learning and
data mining groups. An efficient CNN for arbitrary graph processing was proposed in 2016 [101].
3.6.2. Image Processing and Computer Vision

Most of the models, we have discussed above are applied to different application domains, including
image classification [7–11], detection, segmentation, localization, captioning, video classification and
many more. There is a good survey on DL approaches for image processing and computer vision related
tasks, including image classification, segmentation, and detection [102]. For examples, single image
super-resolution using CNN method [103], image denoising using block-matching CNN [104], photo
aesthetic assessment using A-Lamp (Adaptive Layout-Aware Multi-Patch Deep CNN) [105], DCNN
for hyperspectral imaging segmentation [106], image registration [107], fast artistic style transfer [108],
image background segmentation using DCNN [109], handwritten character recognition [110],
optical image classification [111], crop mapping using high-resolution satellite imagery [112], object
recognition with cellular simultaneous recurrent networks and CNN [113]. The DL approaches are
massively applied to human activity recognition tasks and achieved state-of-the-art performance
Electronics 2019, 8, 292 22 of 66
compared to exiting approaches [114–119]. However, the state-of-the-art models for classification,
segmentation and detection task are listed as follows:
(1) Models for classification problems: according to the architecture of classification models, the
input images are encoded different step with convolution and subsampling layers and finally the
SoftMax approach is used to calculate class probability. Most of the models have discussed above
are applied to the classification problem. However, these model with classification layer can be used
as feature extraction for segmentation and detection tasks. The list of the classification models are
as follows: AlexNet [55], VGGNet [9], GoogleNet [10], ResNet [11], DenseNet [68], FractalNet [69],
CapsuleNet [79], IRCNN [83], IRRCNN [77], DCRN [120] and so on.
(2) Models for segmentation problems: there are several semantic segmentation models have been
proposed in the last few years. The segmentation model consists of two units: Encoding and decoding
units. In the encoding unit, the convolution and subsampling operations are performed to encode to
the lower dimensional latent space where as the decoding unit decodes the image from latent space
performing deconvolution and up-sampling operation. The very first segmentation model is Fully
Convolutional Network (FCN) [27,121]. Later the improved version of this network is proposed which
is named as SegNet [122]. There are several new models have proposed recently which includes
RefineNet [123], PSPNEt [124], DeepLab [125], UNet [126], and R2U-Net [127].
(3) Models for detection problems: the detection problem is a bit different compared to classification
and segmentation problems. In this case, the model goal is to identify target types with its
corresponding position. The model answers two questions: What is the object (classification problem)?
and where the object (regression problem)? To achieve these goals, two losses are calculated for
classification and regression unit in top of the feature extraction module and the model weights
are updated with respect to the both loses. For the very first time, Region based CNN (RCNN) is
proposed for object detection task [128]. Recently, there are some better detection approaches have
been proposed, including focal loss for dense object detector [129], Later the different improved version
of this network is proposed called faster RCNN, fast RCNN [80,130]. mask R-CNN [131], You only
look once (YOLO) [132], SSD: Single Shot MultiBox Detector [133] and UD-Net for tissue detection
from pathological images [120].
3.6.3. Speech Processing

CNNs are also applied to speech processing, such as speech enhancement using multimodal deep
CNN [134], and audio tagging using Convolutional Gated Recurrent Network (CGRN) [135].
3.6.4. CNN for Medical Imaging

Litjens et al [136] provided a good survey on DL for medical image processing, including classification,
detection, and segmentation tasks. Several popular DL methods were developed for medical image
analysis. For instance, MDNet was developed for medical diagnosis using images and corresponding
text description [137], cardiac Segmentation using short-Axis MRI [138], segmentation of optic disc and
retinal vasculature using CNN [139], brain tumor segmentation using random forests with features
learned with fully convolutional neural network (FCNN) [140]. These techniques have been applied in
the field of computational pathology and achieved state-of-the-art performance [28,29,120,141].
4. Advanced Training Techniques

The advanced training techniques or components which need to be considered carefully for
efficient training of DL approach. There are different advanced techniques to apply for training a deep
learning model better. The techniques, including input pre-processing, a better initialization method,
batch normalization, alternative convolutional approaches, advanced activation functions, alternative
pooling techniques, network regularization approaches, and better optimization method for training.
The following sections are discussed on individual advanced training techniques individually.
Electronics 2019, 8, 292 23 of 66
4.1. Preparing Dataset

Presently different approaches have been applied before feeding the data to the network. The
different operations to prepare a dataset are as follows; sample rescaling, mean subtraction, random
cropping, flipping data with respect to the horizon or vertical axis, color jittering, PCA/ZCA whitening
and many more.
4.2. Network Initialization

The initialization of deep networks has a big impact on the overall recognition accuracy [59,60].
Previously, most of the networks have been initialized with random weights. For complex tasks with
high dimensionality data training, a DNN becomes difficult because weights should not be symmetrical
due to the back-propagation process. Therefore, effective initialization techniques are important for
training this type of DNN. However, there are many effective techniques that have been proposed
during the last few years. LeCun [142] and Bengio [143] proposed a simple but effective approach. In
their method, the weights are scaled by the inverse of the square root of the number of input neurons
√
of the layer, which can be stated 1/ Nl , where Nl is the number of input neurons of l th layer. The
deep network initialization approach of Xavier has been proposed based on the symmetric activation
function with respect to the hypothesis of linearity. This approach is known as Xavier initialization
approach. Recently, Dmytro M. et al. [95] proposed Layer-sequential unit-invariance (LSUV), which is
a data-driven initialization approach and provides good recognition accuracy on several benchmark
datasets, including ImageNet. One of the popular initialization approaches has proposed by He et al.
in 2015 [144]. The distribution of the weights of lth layer will be a normal distribution with mean zero
and variance n2 which can be expressed as follows:
l

2
wl ∼ N 0, . (18)
nl
4.3. Batch Normalization

Batch normalization helps accelerate DL processes by reducing internal covariance by shifting
input samples. What that means is the inputs are linearly transformed to have zero mean and unit
variance. For whitened inputs, the network converges faster and shows better regularization during
training, which has an impact on the overall accuracy. Since the data whitening is performed outside
of the network, there is no impact of whitening during training of the model. In the case of deep
recurrent neural networks, the inputs of the nth layer are the combination of n-1th layer, which is
not raw feature inputs. As the training progresses the effect of normalization or whitening reduces
respectively, which causes the vanishing gradient problem. This can slow down the entire training
process and cause saturation. To better training process, batch normalization is then applied to the
internal layers of the deep neural network. This approach ensures faster convergence in theory and
during an experiment on benchmarks. In batch normalization, the features of a layer are independently
normalized with mean zero and variance one [78,145,146]. The algorithm of Batch normalization is
given in Algorithm 1.
Algorithm 1: Batch Normalization (BN)

Inputs: Values of x over a mini-batch: B = { x1,2,3......,m }
Outputs: {yi = BNγ, β (xi )}
µB ← m1 ∑m i=1 xi // mini-batch mean
σB 2
← m1 ∑im=1 ( xi − µB )2 // mini-batch variance
x −µ
x̂i ← qi 2 B // normalize
σ +∈
B
yi = γ x̂i + β ≡ BNγ, β (xi ) // Scaling and shifting
Outputs: y = BN , (x )
μ𝔅 ← ∑ x // mini-batch mean
𝜎𝔅 ← ∑ (𝑥 − 𝜇𝔅 ) // mini-batch variance
𝔅
𝑥 ← // normalize
Electronics 2019, 8, 292 𝔅 ∈ 24 of 66
𝑦 = 𝛾𝑥 + 𝛽 ≡ BN , (x ) // Scaling and shifting

The parameters γ and β are used for the scale and shift factor for the normalization values,
The parameters
so normalization does𝛾 not 𝛽 are
andonly used for
depend on the scale
layer and shift
values. factor
If you use for the normalization
normalization values,the
techniques, so
normalization does not only depend on layer values. If you
following criterions are recommended to consider during implementation: use normalization techniques, the
following criterions are recommended to consider during implementation:
• •Increase the learning
Increase rate rate
the learning
• •Dropout (batch
Dropout normalization
(batch normalizationdoesdoes
the same job) job)
the same
• •L2 weight regularization
L2 weight regularization
• •Accelerating
Accelerating the learning
the learning rate rate decay
decay
• • Remove Local Response Normalization
Remove Local Response Normalization (LRN) (LRN) (if you
(if you usedused
it) it)
• •ShuffleShuffle training
training sample
sample moremore thoroughly
thoroughly
• • Useless distortion of images in the training
Useless distortion of images in the training set set
4.4. Alternative
4.4. Alternative Convolutional
Convolutional Methods
Methods
Alternative and
Alternative and computationally
computationally efficient
efficient convolutional
convolutional techniques
techniques that
that reduce
reduce the
the cost
cost of
of
multiplicationsby
multiplications byaafactor
factorof
of2.5
2.5have
havebeen
beenproposed
proposed[147].
[147].
4.5. Activation
4.5. Activation Function
Function
The traditional
The traditional Sigmoid
Sigmoid and
and Tanh
Tanhactivation
activationfunctions
functionshave
havebeen
beenused
usedfor
for implementing
implementing neural
neural
networkapproaches
network approachesinin thethe past
past fewfew decades.
decades. The graphical
The graphical and mathematical
and mathematical representation
representation is shownis
shown
in in22.
Figure Figure 22.
(a) (b)
Figure 22. Activation

Figure 22. Activation function:
function: (a)
(a) Sigmoid
Sigmoid function,
function, and
and (b)
(b) hyperbolic
hyperbolic transient.
transient.
Sigmoid:
Sigmoid:
1
y= 1 x. (19)
𝑦=1 + e . (19)
1+𝑒
Tanh:
Tanh: e x − e− x
y= . (20)
e𝑒x +−e𝑒− x
𝑦= . (20)
The popular activation function called Rectified 𝑒 + Linear
𝑒 Unit (ReLU) proposed in 2010 solves the
vanishing gradientactivation
The popular problem for training
function deepRectified
called learning Linear
approaches. The basic
Unit (ReLU) conceptin
proposed is 2010
simple to keep
solves the
all the values
vanishing above zero
gradient and sets
problem for all negative
training values
deep to zero
learning that is shown
approaches. Theinbasic
Figure 23 [64].isThe
concept ReLU
simple to
activation
keep
Electronics
was
all the first
2019,values
8, x FOR
used
above in AlexNet [7].
zero and sets all negative values to zero that is shown in Figure 23 [64].
PEER REVIEW The
25 of 67
ReLU activation was first used in AlexNet [7].
Figure 23.
Figure 23. Pictorial
Pictorial representation
representationof
ofRectified
RectifiedLinear
LinearUnit
Unit(ReLU).
(ReLU).
Mathematically we can express ReLU as follows:

𝑦 = max(0, 𝑥). (21)
As the activation function plays a crucial role in learning the weights for deep architectures.
Many researchers focus here because there is much that can be done in this area. Meanwhile, there
are several improved versions of ReLU that have been proposed, which provide even better accuracy
Electronics 2019, 8, 292 25 of 66
Figure 23. Pictorial representation of Rectified Linear Unit (ReLU).

y = max(0, x ). (21)
𝑦 = max(0, 𝑥). (21)
As
Asthe
theactivation
activationfunction
functionplays a crucial
plays rolerole
a crucial in learning the weights
in learning for deep
the weights for architectures. Many
deep architectures.
researchers focus here
Many researchers because
focus there is much
here because therethat can be
is much done
that caninbe
this area.inMeanwhile,
done there are several
this area. Meanwhile, there
improved
are severalversions
improvedof ReLU thatofhave
versions ReLUbeen
thatproposed,
have beenwhich provide
proposed, whicheven better even
provide accuracy
bettercompared
accuracy
to the ReLUtoactivation
compared the ReLUfunction shown
activation in Figure
function shown24. inAn efficient
Figure improved
24. An efficientversion
improved of ReLU
versionactivation
of ReLU
function
activationis function
called theisparametric ReLU (PReLU)
called the parametric ReLUproposed
(PReLU)by KaimingbyHe
proposed et al. inHe
Kaiming 2015.
et al.Figure 25
in 2015.
shows the pictorial representation of Leaky ReLU and ELU activation functions.
Figure 25 shows the pictorial representation of Leaky ReLU and ELU activation functions. This This technique can
automatically
technique canlearn the parameters
automatically learn adaptively and improve
the parameters theand
adaptively accuracy
improveat negligible extraatcomputing
the accuracy negligible
cost
extra[144].
computing cost [144].
(a) (b)
Electronics
Figure2019,
Figure 24.
24.8,Diagram
x FOR PEER
Diagram forREVIEW
for (a)
(a) Leaky
Leaky ReLU
ReLU (Rectified
(RectifiedLinear
LinearUnit),
Unit),and
and(b)
(b) Exponential
ExponentialLinear
LinearUnit 26 of 67
Unit(ELU).
(ELU).
Leaky ReLU:
𝑦 = max(𝑎𝑥, 𝑥). (22)

here 𝑎 is a constant, the value is 0.1.
ELU:
𝑥, 𝑥≥0
𝑦= . (23)
𝑎(𝑒 − 1), 𝑥 < 0
The recent proposal of the Exponential Linear Unit activation function, which allowed for a
faster and more accurate version of the DCNN structure [148]. Furthermore, tuning the negative part
of activation function creates the leaky ReLU with Multiple Exponent Linear Unit (MELU) that are
proposed recently [149]. S shape Rectified Linear Activation units are proposed in 2015 [150]. A
Figure 25. Average and max-pooling
max-pooling operations.
operations.
survey on modern activation functions was conducted in 2015 [151].
Leaky ReLU:Layer or Pooling Layer
4.6. Sub-Sampling
y = max( ax, x ). (22)
At present, two different techniques have been used for the implementation of deep networks
in thea sub-sampling
here is a constant, the value islayer:
or pooling 0.1. Average and max-pooling. The concept of average pooling layer
ELU: for the first time in LeNet [55] and
was used ( AlexNet used Max-pooling layers instead of in 2012[7].
The conceptual diagram for max pooling and x, x ≥pooling
average 0
y= . operation are shown in Figure 25. (23)
The
a(e x − 1), xby<He
concept of special pyramid pooling has been proposed 0 et al. in 2014, which is shown in Figure
26 [152].
The recent proposal of the Exponential Linear Unit activation function, which allowed for a
fasterThe
andmulti-scale pyramid
more accurate pooling
version of the was
DCNNproposed in 2015
structure [148].[153]. In 2015, Benjamin
Furthermore, tuning theG. proposed
negative parta
new
of architecture
activation with creates
function Fractional
themax pooling,
leaky which
ReLU with provides
Multiple state-of-the-art
Exponent Linear classification
Unit (MELU)accuracy
that are
for CIFAR-10 and CIFAR-100 datasets. This structure generalizes the network
proposed recently [149]. S shape Rectified Linear Activation units are proposed in 2015 by considering two
[150]. A survey
important properties for a sub-sampling layer or
on modern activation functions was conducted in 2015 [151].pooling layer. First, the non-overlapped max-
pooling layer limits the generalize of the deep structure of the network, this paper proposed a
network with 3x3 overlapped max-pooling with 2-stride instead of 2×2 as sub-sampling layer [154].
Another paper which has conducted research on different types of pooling approaches, including
mixed, gated, and tree as a generalization of pooling functions [155].
At present, two different techniques have been used for the implementation of deep networks
in the sub-sampling or pooling layer: Average and max-pooling. The concept of average pooling layer
was used for the first time in LeNet [55] and AlexNet used Max-pooling layers instead of in 2012[7].
The conceptual diagram for max pooling and average pooling operation are shown in Figure 25. The
Electronics 2019, 8, 292 26 of 66
concept of special pyramid pooling has been proposed by He et al. in 2014, which is shown in Figure
26 [152].
The multi-scale
4.6. Sub-Sampling pyramid
Layer pooling
or Pooling Layer was proposed in 2015 [153]. In 2015, Benjamin G. proposed a
new architecture with Fractional max pooling, which provides state-of-the-art classification accuracy
At present,
for CIFAR-10 twoCIFAR-100
and different techniques havestructure
datasets. This been usedgeneralizes
for the implementation
the network of bydeep networks
considering twoin
the sub-sampling or pooling layer: Average and max-pooling. The concept of average
important properties for a sub-sampling layer or pooling layer. First, the non-overlapped max- pooling layer
was usedlayer
pooling for the first time
limits in LeNet [55]
the generalize anddeep
of the AlexNet used Max-pooling
structure layers
of the network, instead
this paperofproposed
in 2012 [7].a
The conceptual diagram for max pooling and average pooling operation are
network with 3x3 overlapped max-pooling with 2-stride instead of 2×2 as sub-sampling layer shown in Figure 25.
[154].
The concept
Another paperof special
which haspyramid pooling
conducted has been
research on proposed by Heofetpooling
different types al. in 2014, which is including
approaches, shown in
Figure
mixed,26 [152].and tree as a generalization of pooling functions [155].
gated,
Figure 26.
Figure 26. Spatial
Spatial pyramid
pyramid pooling.
pooling.
4.7. Regularization
The multi-scaleApproaches
pyramidfor DL was proposed in 2015 [153]. In 2015, Benjamin G. proposed a
pooling
new architecture with Fractional max pooling, which provides state-of-the-art classification accuracy
There are different regularization approaches that have been proposed in the past few years for
for CIFAR-10 and CIFAR-100 datasets. This structure generalizes the network by considering two
deep CNN. The simplest but efficient approach called dropout was proposed by Hinton in 2012 [156].
important properties for a sub-sampling layer or pooling layer. First, the non-overlapped max-pooling
In Dropout, a randomly selected subset of activations is set to zero within a layer [157]. The dropout
layer limits the generalize of the deep structure of the network, this paper proposed a network with
concept is shown in Figure 27.
3 × 3 overlapped max-pooling with 2-stride instead of 2 × 2 as sub-sampling layer [154]. Another
paper which has conducted research on different types of pooling approaches, including mixed, gated,
and tree as a generalization of pooling functions [155].
4.7. Regularization Approaches for DL

There are different regularization approaches that have been proposed in the past few years for
deep CNN. The simplest but efficient approach called dropout was proposed by Hinton in 2012 [156].
In Dropout, a randomly selected subset of activations is set to zero within a layer [157]. The dropout
concept
Electronicsis2019,
shown in Figure
8, x FOR 27.
PEER REVIEW 27 of 67
Figure 27.
Figure 27. Pictorial
Pictorial representation
representation of
ofthe
theconcept
conceptDropout.
Dropout.
Another
Another regularization
regularization approach
approach isis called
called Drop
Drop Connect.
Connect. In
In this
this case,
case, instead
instead ofof dropping
dropping the
the
activation, the subset of weights within the network layers are set to zero. As a result,
activation, the subset of weights within the network layers are set to zero. As a result, each layer each layer
receives
receives the
the randomly
randomly selected
selected subset
subset of
of units
units from
from the
theimmediate
immediate previous
previous layer
layer [158].
[158]. Some
Some other
other
regularization
regularizationapproaches
approachesare
areproposed
proposedas aswell
well[159].
[159].
4.8. Optimization Methods for DL

There are different optimization methods, such as SGD, Adagrad, AdaDelta, RMSprop, and
Adam [160]. Some activation functions have been improved upon, such as in the case of SGD where
it was proposed with an added variable momentum, which improved training and testing accuracy.
Electronics 2019, 8, 292 27 of 66
4.8. Optimization Methods for DL

There are different optimization methods, such as SGD, Adagrad, AdaDelta, RMSprop, and
Adam [160]. Some activation functions have been improved upon, such as in the case of SGD where it
was proposed with an added variable momentum, which improved training and testing accuracy. In
the case of Adagrad, the main contribution was to calculate adaptive learning rate during training. For
this method, the summation of the magnitude of the gradient is considered to calculate the adaptive
learning rate. In the case with a large number of epochs, the summation of the magnitude of the
gradient becomes large. The result of this is the learning rate decreases radically, which causes the
gradient to approach zero quickly. The main drawback of this approach is that it causes problems
during training. Later, RMSprop was proposed considering only the magnitude of the gradient of
the immediately previous iteration, which prevents the problems with Adagrad and provides better
performance in some cases. The Adam optimization approach is proposed based on the momentum
and the magnitude of the gradient for calculating adaptive learning rate similar RMSprop. Adam
has improved overall accuracy and helps for efficient training with the better convergence of deep
learning algorithms [161]. The improved version of the Adam optimization approach has been
proposed recently, which is called EVE. EVE provides even better performance with fast and accurate
convergence [162].
5. Recurrent Neural Network (RNN)
5.1. Introduction
Human thoughts have persistence; Human don’t throw a thing away and start their thinking from
scratch in a second. As you are reading this article, you understand each word or sentence based on the
understanding of previous words or sentences. The traditional neural network approaches, including
DNNs and CNNs cannot deal with this type of problem. The standard Neural Networks and CNN
are incapable due to the following reasons. First, these approaches only handle a fixed-size vector
as input (e.g., an image or video frame) and produce a fixed-size vector as output (e.g., probabilities
of different classes). Second, those models operate with a fixed number of computational steps (e.g.,
the number of layers in the model). The RNNs are unique as they allow operation over a sequence of
vectors over time. The Hopfield Newark introduced this concept in 1982 but the idea was described
shortly in 1974 [163]. The pictorial representation is shown in Figure 28.
Figure 28. The

The structure of basic Recurrent Neural Network (RNN) with a loop.
versions of
Different versions ofRNN
RNNhave
havebeen
beenproposed
proposedinin Jordan
Jordan and
and Elman
Elman [164,165].
[164,165]. In the
In the Elman,
Elman, the
the architecture
architecture usesuses
thethe output
output from
from hidden
hidden layers
layers asasinputs
inputsalongside
alongsidethe
thenormal
normalinputs
inputs of
of hidden
layers [129]. On the other hand, the outputs from the output unit are used as inputs with the inputs
of the hidden layer in Jordan network [130]. Jordan, in contrast, uses inputs from the outputs of the
output unit with the inputs to the hidden layer.
layer. Mathematically
Mathematically expressed
expressed as:
as:
Elman network [164]:
[1164]:
ht = σh (wh xt + uh ht−1 + bh ), (24)
h = σ (w x + u h + b ), (24)
y =σ w h +b . (25)
Jordan network [165]:
h = σ (w x + u y + b ), (26)
y =σ w h +b , (27)
architecture uses the output from hidden layers as inputs alongside the normal inputs of hidden
layers [129]. On the other hand, the outputs from the output unit are used as inputs with the inputs
of the hidden layer in Jordan network [130]. Jordan, in contrast, uses inputs from the outputs of the
output unit with the inputs to the hidden layer. Mathematically expressed as:
Elman network [1164]:
Electronics 2019, 8, 292 28 of 66
h = σ (w x + u h + b ), (24)
yyt =
= σσy w
wy h
ht +
+bby .

(25)
(25)
Jordan network [165]:

ht = σh (w
wh xt + uh yt−1 + bh),, (26)
h =σ x +u y +b (26)

yyt = σy w y ht +
= σ w h + b ,, b y (27)
(27)
where
where xxt isisa avector
vectorofof
inputs, ht are
inputs, hidden
h are layerlayer
hidden vectors, yt areytheare
vectors, output vectors,
the output w and w
vectors, u are
andweight
u are
matrices and b is the bias vector.
weight matrices and b is the bias vector.
A loop
A loop allows
allows information
information to to be
be passed
passed from
from one
one step
step of
of the
the network
network toto the
the next.
next. AA recurrent
recurrent
neural network can be thought of as multiple copies of the same network, each network passing aa
neural network can be thought of as multiple copies of the same network, each network passing
message to
message to aa successor.
successor. The
The diagram
diagram below
below Figure
Figure 29
29 shows
shows what
what happens
happens ifif we
we unroll
unroll the
the loop.
loop.
Figure 29. An unrolled RNNs.
The main problem with RNN approaches is that there exists the vanishing vanishing gradient
gradient problem.
problem.
For the first time, this problem is solved by Hochreiter et al. [166]. [166]. A deep RNN consisting of 1000
subsequent
Electronics 2019,layers
8, x FORwas
PEERimplemented
REVIEW and evaluated to solve deep learning tasks in 1993 [167].29 There
of 67
are several solutions that have been proposed for solving the vanishing gradient problem of RNN
approaches
are in the pastunderstanding
used for language few decades. [170].
Two possible effectivemodeling,
In the language solutions ittotries
thisto
problem
predict are
the first
nextto clip
word
or
theset of words
gradient andorscale
some cases
the if the
sentences
gradient if thebased
norm is
norm is too
on too large,
thelarge, and secondly,
previous
and secondly,
ones create
[171].create
RNNs aa better
better RNN model.
are networks
RNN model.
with
One of the better models was introduced by
loops in them, allowing information to persist. Another Felix A. et al. in 2000 named Long Short-Term
el at. example: The RNNs are able to connect Memory
(LSTM)
(LSTM) [168,169].
previous information
[168,169]. From
Fromtothe
theLSTM
present
LSTM there have
task:
there been
Using
have been different
previous
different advanced
video approaches
frames,
advanced proposed
understanding
approaches theinpresent
proposed the
in last
the
few trying
and
last years which
few years are explained
to generate
which future
are in the as
frames
explained following
in well sections.sections.
[172].
the following The diagram for LSTMfor
The diagram is shown
LSTM is in shown
Figure 30.
in
Figure 30.
The RNN approaches allowed sequences in the input, the output, or in the most general case
both. For example, DL for text mining, building deep learning models on textual data requires
representation of the basic text unit and word. Neural network structures that can hierarchically
capture the sequential nature of the text. In most of these cases, RNNs or Recursive Neural Networks
Figure 30.
Figure Diagramfor
30. Diagram forLong
LongShort-Term
Short-Term Memory
Memory (LSTM).
The Short-Term
5.2. Long RNN approachesMemoryallowed
(LSTM) sequences in the input, the output, or in the most general case both.
For example, DL for text mining, building deep learning models on textual data requires representation
of theThe keytext
basic idea of and
unit LSTMs is the
word. cell state,
Neural networkthestructures
horizontalthat
linecan
running throughcapture
hierarchically the toptheof Figure 31.
sequential
LSTMs remove
nature of or In
the text. add information
most to theRNNs
of these cases, cell state called gates:
or Recursive An Networks
Neural (i used
input gateare ), forget gate (f )
for language
and output gate (o ) can be defined as:
understanding [170]. In the language modeling, it tries to predict the next word or set of words or some
cases sentences based on the previousf ones = σ(W [171]. ,x +
. h RNNs are ),
b networks with loops in them, allowing
(28)
information to persist. Another example: The RNNs are able to connect previous information to the
i = σ(W . h , x + b ), (29)
present task: Using previous video frames, understanding the present and trying to generate future
frames as well [172]. C = tanh(W . h , x + b ), (30)
C =f ∗C +i ∗ C , (31)
O = σ(W . h , x + b ), (32)
h = O ∗ tanh(C ). (33)
Electronics 2019, 8, 292 29 of 66
Figure 30. Diagram for Long Short-Term Memory (LSTM).

5.2. Long Short-Term Memory (LSTM)
5.2. Long Short-Term Memory (LSTM)
The key idea of LSTMs is the cell state, the horizontal line running through the top of Figure 31.
LSTMs The key idea
remove of LSTMs
or add is the cell
information state,
to the cellthe horizontal
state line An
called gates: running
input through
gate (it ),the top gate
forget of Figure 31.
(ft ) and
LSTMs remove or add information
output gate (ot ) can be defined as: to the cell state called gates: An input gate (i ), forget gate (f )
and output gate (o ) can be defined as:
ft = σ(Wf ·[ht−1 , xt ] + bf ), (28)
f = σ(W . h , x + b ), (28)
(Wi ·[. hht−1 ,,xxt ] +
iti == σσ(W + bbi )),, (29)
(29)
e tC==tan
C h(WC ·[. hhC−1 , x t ] + b C),
tanh(W ), (30)
(30)
CCt ==ftf ∗∗CCt−1 + it ∗ C
eCt ,, (31)
(31)
OO
t == σσ(W + bbO )),,
(WO ·[. hht−1 ,, xxt ] + (32)
(32)
hht ==OOt ∗∗tan h(Ct ))..

tanh(C (33)
(33)
LSTM models
LSTM models are
arepopular
popularforfortemporal
temporalinformation
informationprocessing.
processing. Most
Most of
of the
the papers
papers that
that include
include
LSTM models with some minor variance. Some of them are discussed in the following
LSTM models with some minor variance. Some of them are discussed in the following section. Theresection. There
is aaslightly
is slightly modified
modified version
version of
ofthe
thenetwork
networkwith
withpeephole
peepholeconnections
connections by
by Gers
GersandandSchimidhuber
Schimidhuber
proposedin
proposed in2000
2000[168].
[168]. The
The concept
concept of ofpeepholes
peepholesisisincluded
includedwith
withalmost
almostall
allthe
thegated
gatedin
inthis
thismodel.
model.
Figure 31. Diagram

Figure 31. Diagram for
for Gated
Gated Recurrent
RecurrentUnit
Unit(GRU).
(GRU).
5.3. Gated Recurrent Unit (GRU)

5.3. Gated Recurrent Unit (GRU)
GRU also came from LSTMs with slightly more variation [173]. GRUs are now popular in the
GRU also came from LSTMs with slightly more variation [173]. GRUs are now popular in the
community who are working with recurrent networks. The main reason for the popularity is the
community who are working with recurrent networks. The main reason for the popularity is the
computation cost and simplicity of the model, which is shown in Figure 31. GRUs are lighter versions
computation cost and simplicity of the model, which is shown in Figure 31. GRUs are lighter versions
of RNN approaches than standard LSTM in term of topology, computation cost and complexity [173].
This technique combines the forget and input gates into a single update gate and merges the cell state
and hidden state along with some other changes. The simpler model of the GRU has been growing
increasingly popular. Mathematically the GRU can be expressed with the following equations:
zt = σ(Wz ·[ht−1 , xt ]), (34)
rt = σ(Wr ·[ht−1 , xt ]), (35)

et = tan h(W·[rt ∗ ht−1 , xt ]),
h (36)
ht = (1 − zt ) ∗ ht−1 + zt ∗ h
et . (37)
The question is which one is the best? According to the different empirical studies, there is no
clear evidence of a winner. However, the GRU requires fewer network parameters, which makes the
model faster. On the other hand, LSTM provides better performance, if you have enough data and
computational power [174]. There is a variant LSTM named Deep LSTM [175]. Another variant that is
h = tanh(W. r ∗ h , x ), (36)
h = (1 − z ) ∗ h +z ∗ h . (37)
The question is which one is the best? According to the different empirical studies, there is no
clear evidence of a winner. However, the GRU requires fewer network parameters, which makes the
Electronics 2019, 8, 292 30 of 66
model faster. On the other hand, LSTM provides better performance, if you have enough data and
computational power [174]. There is a variant LSTM named Deep LSTM [175]. Another variant that
a bit
is adifferent approach
bit different approachcalled A clockwork
called A clockworkRNN RNN[176]. There
[176]. is an
There important
is an important empirical
empiricalevaluation
evaluation
onon a different
a different version
version ofof
RNN
RNN approaches,
approaches,including
including LSTM
LSTM byby
Greff, et et
Greff, al.al.
inin
2015
2015[177] and
[177] andthe final
the final
conclusion
conclusionwas wasallallthe
theLSTM
LSTMvariants
variantswere
wereallallabout
aboutthe
thesame
same [177].
[177]. Another empirical evaluation
evaluation is
is conducted
conducted on on thousands of RNN RNN architecture,
architecture, including
includingLSTM,
LSTM,GRUGRUand andsosoon
onfinding
findingsome
somethat
that
worked better than LSTMs on certain tasks
worked better than LSTMs on certain tasks [178] [178]
5.4. Convolutional
5.4. LSTM
Convolutional (ConvLSTM)
LSTM (ConvLSTM)
The
Theproblem
problemwithwithfully
fullyconnected
connected(FC) (FC)LSTM
LSTMand andshort
shortFC-LSTM
FC-LSTMmodelmodelis ishandling
handling
spatiotemporal
spatiotemporal data and and
data its usage of full connections
its usage in the input-to-state
of full connections and state-to-state
in the input-to-state and transactions,
state-to-state
where no spatialwhere
transactions, information has been
no spatial encoded.has
information Thebeen
internal gates The
encoded. of ConvLSTM are 3D
internal gates of tensors,
ConvLSTM whereare
the
3D tensors, where the last two dimensions are spatial dimensions (rows and columns). the
last two dimensions are spatial dimensions (rows and columns). The ConvLSTM determines The
future state ofdetermines
ConvLSTM a certain cell infuture
the the grid with
state of respect
a certaintocell
inputs and
in the thewith
grid pastrespect
states of
toits local and
inputs neighbors
the past
which can be achieved using convolution operations in the state-to-state or inputs-to-states
states of its local neighbors which can be achieved using convolution operations in the state-to-state transition,
shown in Figure 32. transition, shown in Figure 32.
or inputs-to-states
Figure
Figure 32.32. Pictorial
Pictorial diagram
diagram forfor ConvLSTM.
ConvLSTM.
ConvLSTM
ConvLSTM is is
providing good
providing good performance
performanceforfor
temporal
temporaldata
dataanalysis
analysiswith
withvideo
videodatasets
datasets[172].
[172].
Mathematically
MathematicallythetheConvLSTM
ConvLSTM is is
expressed asas
expressed follows
follows where
where* represents
* representsthe convolution
the convolutionoperation
operation
and ◦ denotes
and ∘ denotes
forfor
Hadamard
Hadamardproduct:
product:
it =i σ=(w
σ(w· X.t𝒳
xi ++ w∗ H
whi ∗ ℋ ++
t−1 ww ∘ 𝒞 + b ),
hi ◦ Ct−1 + bi ),
(38)
(38)
f = σ(w . 𝒳 + w ∗ ℋ + w ∘ 𝒞 + b ), (39)
ft = σ(wxf · Xt + whf ∗ Ht−1 + whf ◦ Ct−1 + bf ), (39)
C = tanh(w . 𝒳 + w ∗ ℋ + b ), (40)
Cet = tan h(wxc · Xt + whc ∗ Ht−1 + bC ), (40)
C =f ∘C +i ∗C , (41)
Ct = ft ◦ Ct−1 + it ∗ Cet , (41)
o = σ(w . 𝒳 + w ∗ ℋ + w ∘ 𝒞 + b , (42)
ot = σ(wxo · Xt + who ∗ Ht−1 + who ◦ Ct + bo , (42)
h = o ∘ tanh(C ). (43)
ht = ot ◦ tan h(Ct ). (43)
5.5.
5.5. AA variantofofArchitectures
Variant ArchitecturesofofRNN
RNNwith
withRespective
RespectivetotothetheApplications
Applications
To incorporate the attention mechanism with RNNs, Word2Vec is used in most of the cases for
a word or sentence encoding. Word2vec is a powerful word embedding technique with a 2-layer
predictive NN from raw text inputs. This approach is used in the different fields of applications,
including unsupervised learning with words, relationship learning between the different words, the
ability to abstract higher meaning of the words based on the similarity, sentence modeling, language
understanding and many more. There are different other word embedding approaches that have been
proposed in the past few years which are used to solve difficult tasks and provide state-of-the-art
performance, including machine translation and language modeling, Image and video captioning and
time series data analysis [179–181].
From the application point of view, RNNs can solve different types of problems which need
different architectures of RNNs, shown in Figure 33. In Figure 33, Input vectors are represented as
ability to abstract higher meaning of the words based on the similarity, sentence modeling, language
understanding and many more. There are different other word embedding approaches that have
been proposed in the past few years which are used to solve difficult tasks and provide state-of-the-
art performance, including machine translation and language modeling, Image and video captioning
and time2019,
Electronics series data analysis [179–181].
8, 292 31 of 66
From the application point of view, RNNs can solve different types of problems which need
different architectures of RNNs, shown in Figure 33. In Figure 33, Input vectors are represented as
green, RNN
green, RNNstatesstatesareare
represented
representedwithwith
blue blue
and orange represents
and orange the output
represents thevector.
outputThese structures
vector. These
can be described as:
structures can be described as:
One to
One to One: Standard mode
One: Standard mode forfor classification
classification without
without RNN RNN (e.g.,
(e.g., image
image classification
classification problem)
problem)
shown Figure
shown Figure 33(a) 33a.
Many to
Many to One: Sequence of
One: Sequence of inputs
inputs and
and aa single
single output
output (e.g.,
(e.g., the
the sentiment
sentiment analysis
analysis where
where inputs
inputs
are aa set
are set of
of sentences
sentences or or words
words andand output
output isis aa positive
positive or or negative
negative expression)
expression) shown
shown Figure
Figure 33b.
33b.
One to Many: Where a system takes an input and produces a sequence of outputs (Image
One to Many: Where a system takes an input and produces a sequence of outputs (Image
Captioning problem:
Captioning problem:InputInputis is a single
a single image
image andand output
output is aofset
is a set of words
words with context)
with context) shown shown
Figure
Figure
33c. 33c.
Many to
Many Many: Sequences
to Many: Sequences of of inputs
inputs andand outputs
outputs (e.g.,
(e.g., machine
machine translation: machine takes
translation: machine takes a a
sequence of
sequence of words
words from
from English
English andand translates
translates to to aa sequence
sequence of of words
words in in French)
French) shown
shown Figure
Figure33d.
33d.
Many to Many: Sequence to sequence learning (e.g., video classification problem in which
Many to Many: Sequence to sequence learning (e.g., video classification problem in which we we
take video frames as input and wish to label each frame of the video
take video frames as input and wish to label each frame of the video shown Figure 33e. shown Figure 33e.
(a) (b) (c)
(d) (e)
Figure 33. The different structure of RNN with respect to the applications: (a) One to one; (b) many to
Figure
one; (c) 33.
oneThe different
to many; (d) structure of RNN
many to many; with
and respecttotomany.
(e) many the applications: (a) One to one; (b) many
to one; (c) one to many; (d) many to many; and (e) many to many.
5.6. Attention-based Models with RNN
Different attention-based models have been proposed using RNN approaches. The first initiative
for RNNs with the attention that automatically learns to describe the content of images is proposed
by Xu, et al. in 2015 [182]. A dual state attention based RNN is proposed for effective time series
prediction [183]. Another difficult task is Visual Question Answering (VQA) using GRUs where the
inputs are an image and a natural language question about the image, the task is to provide an accurate
natural language answer. The output is to be conditional on both image and textual inputs. A CNN
is used to encode the image and an RNN is implemented to encode the sentence [184]. Another
outstanding concept is released from Google called Pixel Recurrent Neural Networks (Pixel RNN).
This approach provides state-of-the-art performance for image completion tasks [185]. The new model
called residual RNN is proposed, where the RNN is introduced with an effective residual connection
in a deep recurrent network [186].
Electronics 2019, 8, 292 32 of 66
5.7. RNN Applications

RNNs, including LSTM and GRU, are applied to Tensor processing [187]. Natural Language
Processing using RNN techniques, including LSTMs and GRUs [188,189]. Convolutional RNNs based
on multi-language identification system has been proposed in 2017 [190]. Time series data analysis
using RNNs [191]. Recently, TimeNet was proposed based on pre-trained deep RNNs for time series
classification (TSC) [192]. Speech and audio processing, including LSTMs for large-scale acoustic
modeling [193,194]. Sound event prediction using convolutional RNNs [195]. Audio tagging using
Convolutional GRUs [196]. Early heart failure detection is proposed using RNNs [197].
RNNs are applied in tracking and monitoring: Data-driven traffic forecasting systems are
proposed using Graph Convolutional RNN (GCRNN) [25]. An LSTM based network traffic prediction
system is proposed with a neural network-based model [198]. Bidirectional Deep RNN is applied
to driver action prediction [199]. Vehicle Trajectory prediction using an RNN [200]. Action
recognition using an RNN with a Bag-of-Words [201]. Collection anomaly detection using LSTMs for
cybersecurity [202].
6. Auto-Encoder (AE) and Restricted Boltzmann Machine (RBM)

This section will be discussing one of the unsupervised deep learning approaches the Auto
Encoder [61] (e.g., variational auto-encoder (VAE) [203], denoising AE [65], sparse AE [204], stacked
denoising AE [205], Split-Brain AE [206]). The applications of different AE are also discussed at the
end of this chapter.
6.1. Review of Auto-Encoder (AE)

An AE is a deep neural network approach used for unsupervised feature learning with efficient
data encoding and decoding. The main objective of autoencoder is to learn and represent (encoding) of
the input data, typically for data dimensionality reduction, compression, fusion and many more. This
autoencoder technique consists of two parts: The encoder and the decoder. In the encoding phase, the
input samples are mapped usually in the lower dimensional features space with a constructive feature
representation. This approach can be repeated until the desired feature dimensional space is reached.
Whereas in the decoding phase, we regenerate actual features from lower dimensional features with
reverse processing. The conceptual diagram of auto-encoder with encoding and decoding phases is
shown in2019,
Electronics Figure 34. PEER REVIEW
8, x FOR 33 of 67
Figure 34. Diagram

Figure 34. Diagram for
for Auto
Auto encoder.
encoder.
The encoder and decoder transition can be represented with ∅. and ϕ, ∅ : X → F and
The encoder and decoder transition can be represented with ∅ and 𝜑, ∅ ∶ 𝒳 → ℱ and 𝜑 ∶ ℱ →
𝒳, then→ X , then
ϕ : F
∅, ϕ = argmin∅, ϕ k X − (∅, ϕ) X k2 . (44)
∅, 𝜑 = 𝑎𝑟𝑔𝑚𝑖𝑛∅, ‖𝑋 − (∅, 𝜑)𝑋‖ . (44)
If we consider a simple autoencoder with one hidden layer, where the input is 𝑥 ∈ ℝ = 𝒳,
which is mapped onto ∈ ℝ = ℱ, it can be then expressed as follows:
𝑧 = 𝜎 (𝑊𝑥 + 𝑏), (45)
where W is the weight matrix and b is bias. 𝜎 represents an element wise activation function, such
Electronics 2019, 8, 292 33 of 66
If we consider a simple autoencoder with one hidden layer, where the input is x ∈ Rd = X , which
is mapped onto ∈ R p = F , it can be then expressed as follows:
z = σ1 (Wx + b), (45)
where W is the weight matrix and b is bias. σ1 represents an element wise activation function, such as
a sigmoid or a rectified linear unit (RLU). Let us consider z is again mapped or reconstructed onto x 0
which is the same dimension of x. The reconstruction can be expressed as
x 0 = σ2 W 0 z + b0 .

(46)
This model is trained with minimizing the reconstruction errors, which is defined as loss function
as follows
2 2
L x, x 0 = k x − x 0 k = k x − σ2 W 0 (σ1 (Wx + b)) + b0 k .

(47)
Usually, the feature space of F has lower dimensions than the input feature space X , which can
be considered as the compressed representation of the input sample. In the case of multilayer auto
encoder, the same operation will be repeated as required with in the encoding and decoding phases.
A deep Auto encoder is constructed by extending the encoder and decoder with multiple hidden
layers. The Gradient vanishing problem is still a big issue with the deeper model of AE: The gradient
becomes too small as it passes back through many layers of an AE model. Different advanced AE
models are discussed in the following sections.
6.2. Variational Autoencoders (VAEs)

There are some limitations of using simple Generative Adversarial Networks (GAN) which are
discussed in Section 7. At first, images are generated using GAN from input noise. If someone wants to
generate a specific image, then it is difficult to select the specific features (noise) randomly to produce
desired images. It requires searching the entire distribution. Second, GANs differentiate between ‘real’
and ‘fake’ objects. For example, if you want to generate a dog, there is no constraint that the dog must
look like a dog. Therefore, it produces same style images which the style looks like a dog but if we
closely observed then it is not exactly. However, VAE is proposed to overcome those limitations of
basic GANs, where the latent vector space is used to represent the images which follow a unit Gaussian
distribution [203,207]. The conceptual diagram for VAE is shown in Figure 35.
Figure 35. Variational Auto-Encoder.

Figure 35. Variational Auto-Encoder.
In this model, there are two losses, one is a mean squared error that determines, how good the
In this model, there are two losses, one is a mean squared error that determines, how good the
network is doing for reconstructing the image, and loss (the Kullback-Leibler (KL) divergence) of
network is doing for reconstructing the image, and loss (the Kullback-Leibler (KL) divergence) of
latent, which determines how closely the latent variable match is with unit Gaussian distribution. For
example, suppose x is an input and the hidden representation is z. The parameters (weights and biases)
example, suppose 𝑥 is an input and the hidden representation is z. The parameters (weights and
are θ. For reconstructing the phase the input is z and the desired output is x. The parameters (weights
biases) are 𝜃 . For reconstructing the phase the input is 𝑧 and the desired output is 𝑥 . The
and biases) are φ. So, we can represent the encoder as qθ (z| x ) and decoder pφ ( x |z) respectively. The
parameters (weights and biases) are 𝜙. So, we can represent the encoder as 𝑞 (𝑧|𝑥) and decoder
loss function of both networks and latent space can be represented as
𝑝 (𝑥|𝑧) respectively. The loss function of both networks and latent space can be represented as
𝑙 (𝜃, 𝜙) = −𝐸 ~ ( | ) 𝑙𝑜𝑔𝑝 (𝑥 |𝑧) + 𝐾𝐿 𝑞 (𝑧|𝑥 )| 𝑝(𝑧) . (48)
6.3. Split-Brain Autoencoder

Recently Split-Brain AE was proposed from Berkeley AI Research (BAIR) lab, which is the
example, suppose 𝑥 is an input and the hidden representation is z. The parameters (weights and
biases) are 𝜃 . For reconstructing the phase the input is 𝑧 and the desired output is 𝑥 . The
parameters (weights and biases) are 𝜙. So, we can represent the encoder as 𝑞 (𝑧|𝑥) and decoder
𝑝 (𝑥|𝑧) 2019,
Electronics respectively.
8, 292 The loss function of both networks and latent space can be represented as
34 of 66
𝑙 (𝜃, 𝜙) = −𝐸 ~ ( | ) 𝑙𝑜𝑔𝑝 (𝑥 |𝑧) + 𝐾𝐿 𝑞 (𝑧|𝑥 )| 𝑝(𝑧) . (48)

li (θ, φ) = − Ez∼qθ (z| xi ) logpφ ( xi |z) + KL(qθ (z| xi )| p(z)). (48)
Recently Split-Brain AE was proposed from Berkeley AI Research (BAIR) lab, which is the
Recently Split-Brain AE
architectural modification was proposed
of traditional from Berkeley
autoencoders AI Researchrepresentation
for unsupervised (BAIR) lab, which is the
learning. In
architectural modification
this architecture, the networkof traditional autoencoders
is split into for unsupervised
disjoint sub-networks, representation
where two learning.
networks try In
to predict
this
the architecture, the network
feature representation of is
ansplit into
entire disjoint
image sub-networks,
[206]. where
The following two36
Figure networks try concept
shows the to predictof
the feature representation
split-Brain Autoencoder. of an entire image [206]. The following Figure 36 shows the concept of
split-Brain Autoencoder.
Figure36.
Figure 36.Split-Brain
Split-BrainAutoencoder.
Autoencoder.
6.4.
6.4.Applications
ApplicationsofofAE
AE
AE
AE isis applied
applied inin Bio-informatics
Bio-informatics [136,208]
[136,208] and
andcybersecurity
cybersecurity [209].
[209]. We
We can
can apply
apply AE
AE forfor
unsupervised feature extraction and then apply Winner Take All (WTA) for clustering those
unsupervised feature extraction and then apply Winner Take All (WTA) for clustering those samples samples
for
forgenerating
generatinglabels
labels[210]. AEAE
[210]. hashas
been used
been as an
used as encoding and decoding
an encoding technique
and decoding with or
technique for or
with other
for
deep learning approaches, including CNN, DNN, RNN, and RL in the last decade. However,
other deep learning approaches, including CNN, DNN, RNN, and RL in the last decade. However, here are
some other
here are someapproaches recently published
other approaches [207,211]. [207,211]
recently published
6.5. Review of RBM

6.5. Review of RBM
Restricted Boltzmann Machine (RBM) is another unsupervised deep learning approach. The
Restricted Boltzmann Machine (RBM) is another unsupervised deep learning approach. The
training phase can be modeled using a two-layer network called a Restricted Boltzmann Machine [212]
training phase can be modeled using a two-layer network called a Restricted Boltzmann Machine
in which stochastic binary pixels are connected to stochastic binary feature detectors using
[212] in which stochastic binary pixels are connected to stochastic binary feature detectors using
symmetrically weighted connections. RBM is an energy-based undirected generative model that
symmetrically weighted connections. RBM is an energy-based undirected generative model that uses
uses a layer of8,hidden variables to model distribution over visible variables. The undirected model for
aElectronics
layer of2019,
hidden x FOR PEER
variablesREVIEW
to model distribution over visible variables. The undirected model for35 of 67
the
the interactions between the hidden and visible variables is used to ensure that the contribution of
interactions between the hidden and visible variables is used to ensure that the contribution of the
likelihood
the term
likelihood termto to
thethe
posterior over
posterior overthe
thehidden
hiddenvariables
variablesisisapproximately
approximately factorial which
which greatly
greatly
facilitatesinference
facilitates inference[213].
[213].The
Theconceptual
conceptualdiagram
diagramof
ofRBM
RBMisisshown
shownininFigure
Figure37.
37.
Figure 37.Block
Figure37. Blockdiagram
diagramfor
forRestricted
RestrictedBoltzmann
BoltzmannMachine
Machine(RBM).
(RBM).
Energy-based models mean that the probability distribution over the variables of interest is
defined through an energy function. The energy function is composed from a set of observable
variables s 𝑉 = {𝑣 } and a set of hidden variables = {ℎ } , where i is a node in the visible layer, j is a
node in the hidden layer. It is restricted in the sense that there are no visible-visible or hidden-hidden
connections. The values corresponding to visible units of the RBM because their states are observed;
the feature detectors correspond to hidden units. A joint configuration, (v,h) of the visible and hidden
Electronics 2019, 8, 292 35 of 66
Energy-based models mean that the probability distribution over the variables of interest is
defined through an energy function. The energy function is composed from a set of observable
variables s V = {vi } and a set of hidden variables = {hi }, where i is a node in the visible layer, j is a
node in the hidden layer. It is restricted in the sense that there are no visible-visible or hidden-hidden
connections. The values corresponding to visible units of the RBM because their states are observed;
the feature detectors correspond to hidden units. A joint configuration, (v,h) of the visible and hidden
units has an energy (Hopfield, 1982) given by:
E(v, h) = − ∑ ai vi − ∑ b j h j − ∑ ∑ vi wi,j h j , (49)

i j i j
where vi h j are the binary states of visible unit i and hidden unit j, ai , b j are their biases and wij is the
weight between them. The network assigns a probability to a possible pair of a visible and a hidden
vector via this energy function:
1
p(v, h) = e−E(v,h) , (50)
Z
where the partition function, Z, is given by summing over all possible pairs of visible and hidden vectors:
Z = ∑ e−E(v,h) . (51)
v,h
The probability that the network assigns to a visible vector, v, is given by summing over all
possible hidden vectors:
1
p(v) = ∑ e−E(v,h) . (52)
Z h
The probability that the network assigns to a training sample can be raised by adjusting the
weights and biases to lower the energy of that sample and to raise the energy of other samples,
especially those have low energies and therefore make a big contribution to the partition function. The
derivative of the log probability of a training vector with respect to weight is surprisingly simple.
∂logp(v)

= vi h j data − vi h j model , (53)
∂wij
where the angle brackets are used to denote expectations under the distribution specified by the
subscript that follows. This leads to a simple learning rule for performing stochastic steepest ascent in
the log probability of the training data:

wij = ε vi h j data
− v i h j model
, (54)
where ε is a learning rate. Given a randomly selected training image, v, the binary state, h j , of each
hidden unit, j is set to 1 with probability
!
p h j = 1|v = σ bj + ∑ vi wij ,

(55)
i

where σ(x) is the logistic sigmoid function 1/ 1 + e(−x) , vi h j is then an unbiased sample. Because
there are no direct connections between visible units in an RBM, it is also easy to get an unbiased
sample of the state of a visible unit, given a hidden vector
!
p(vi = 1|h) = σ ai + ∑ h j wij . (56)
j
Electronics 2019, 8, 292 36 of 66

Getting an unbiased sample of vi h j model is much more difficult. It can be done by starting at any
random state of the visible units and performing alternating Gibbs sampling for a long time. A single
iteration of alternating Gibbs sampling consists of updating all the hidden units in parallel using
Equation (55) followed by updating all the visible units in parallel using the following Equation (56).
A much faster learning procedure was proposed in Hinton (2002). This starts by setting the states of
the visible units to a training vector. Then the binary states of the hidden units are all computed in
parallel using Equation (55). Once binary states have been chosen for the hidden units, a reconstruction
is produced by setting each vi to 1 with a probability given by Equation (56). The change in weight is
then given by

∆wij = ε vi h j data − vi h j recon .

(57)
A simplified version of the same learning rule that uses the states of individual units instead
of a pairwise product is used for the biases [214]. This approach is mainly used for pre-training a
neural network in an unsupervised manner to generate initial weights. One of the most popular deep
learning approaches called Deep Belief Network (DBN) is proposed based on this approach. Some
of the examples of the applications with RBM and DBN for data encoding, news clustering, image
segmentation, and cybersecurity are shown, for detail see References [57,215–217].
7. Generative Adversarial Networks (GAN)

At the beginning of this chapter, we started with a quote from Yann LeCun, GAN is the best
concept proposed in the last ten years in the field of deep learning (Neural networks).
7.1. Review on GAN

The concept of generative models in machine learning started a long time before which is used
for data modeling with conditional probability density function. Generally, this type of model is
considered a probabilistic model with a joint probability distribution over observation and target
(label) values. However, we did not see the big success of this generative model before. Recently, deep
learning-based generative models have become popular and shown enormous success in different
application domains.
Deep learning is a data-driven technique that performs better as the number of input samples
increased. Due to this reason, learning with reusable feature representations from a huge number
of the un-labels dataset has become an active research area. We mentioned in the introduction that
Computer vision has different tasks, segmentation, classification, and detection, which requires large
amounts of labeled data. This problem has been attempted to be solved be generating similar samples
with a generative model.
Generative Adversarial Network (GAN) is a deep learning approach recently invented by
Goodfellow in 2014. GANs offer an alternative approach to maximum likelihood estimation techniques.
GAN is an unsupervised deep learning approach where two neural networks compete against each
other in a zero-sum game. In the case of the image generation problem, the generator starts with
Gaussian noise to generate images and the discriminator determines how good the generated images
are. This process continues until the outputs of the generator become close to actual input samples.
According to Figure 38, it can be considered that Discriminator (D) and Generator (G) two players
playing the min-max game with the function of V (D, G) which can be expressed as follows according
to this paper [33,218].
minG maxD V (D, G) = Ex∼Pdata (x) [log(D(x))] + Ez∼Pdata (z) [log(1 − D(G(z)))]. (58)
In practice, this equation may not provide sufficient gradient for learning G (which started from
random Gaussian noise) at the early stages. In the early stages, D can reject samples because they
are clearly different compared to training samples. In this case, log(1 − D(G(z))) will be saturated.
Instead of training G to minimize log(1 − D(G(z))) we can train G to maximize log(G(z)) objective
techniques.
Electronics 2019,GAN
8, 292 is an unsupervised deep learning approach where two neural networks compete 37 of 66
against each other in a zero-sum game. In the case of the image generation problem, the generator
starts with Gaussian noise to generate images and the discriminator determines how good the
function which
generated images provides
are. Thismuch better
process gradients
continues in the
until early stagesof
outputs during learning.become
the generator However, there
close were
to actual
some limitations of convergence during training with the first version. In the beginning
input samples. According to Figure 38, it can be considered that Discriminator (D) and Generator (G) state a GAN
has some
two players limitations
playing the regarding
min-maxthe game
following
withissues:
the function of V (D, G) which can be expressed as
follows
• Theaccording to this paper
lack of a heuristic cost [33,218].
function (as pixel-wise approximate means square errors (MSE))
• 𝑚𝑖𝑛
Unstable 𝑚𝑎𝑥 (sometimes
to train 𝑉(𝐷, 𝐺) = 𝔼that
~ ) 𝑙𝑜𝑔(𝐷(𝑥)) + 𝔼 ~
can( because ( ) 𝑙𝑜𝑔(1 − 𝐷(𝐺(𝑧))) .
of producing nonsensical outputs) (58)
Figure 38. Conceptual diagram for Generative Adversarial Networks (GAN).

Figure 38. Conceptual diagram for Generative Adversarial Networks (GAN).
Research in the area of GANs has been ongoing with many improved versions being
In practice, this equation may not provide sufficient gradient for learning G (which started from
proposed [219]. GANs are able to produce photorealistic images for applications, such as visualization
random Gaussian noise) at the early stages. In the early stages, D can reject samples because they are
of interior or industrial design, shoes, bags, and clothing items. GAN is also extensively used in
clearly different compared to training samples. In this case, 𝑙𝑜𝑔(1 − 𝐷(𝐺(𝑧))) will be saturated.
the field of game development and artificial video generation [220]. GANs have two different areas
Instead of training G to minimize 𝑙𝑜𝑔(1 − 𝐷(𝐺(𝑧))) we can train G to maximize 𝑙𝑜𝑔(𝐺(𝑧)) objective
of DL that they fall into semi-supervised and unsupervised. Some research in these areas focuses
function which provides much better gradients in early stages during learning. However, there were
on the topology of the GAN architecture to improve functionality and the training approach. Deep
some limitations of convergence during training with the first version. In the beginning state a GAN
convolution GAN (DCGAN) is a convolution-based GAN approach proposed in 2015 [221]. This
has some limitations regarding the following issues:
semi-supervised approach has shown promised results compared to its unsupervised counterpart. The
• The lack of a heuristic cost function (as pixel-wise approximate means square errors
regenerated results from DCGAN have shown in the following figures [183]. Figure 39 according to
(MSE))
article in [221], shows the output for generated bedroom images after one training pass through the
• Unstable to train (sometimes that can because of producing nonsensical outputs)
dataset. Most of the figures included in this section are generated through experiments. Theoretically,
Research in the area of GANs has been ongoing with many improved versions being proposed
the model could learn to memorize training examples, but this is experimentally unlikely as we train
[219]. GANs are able to produce photorealistic images for applications, such as visualization of
with a small learning rate and mini batches with SGD. We are aware of no prior empirical evidence
interior or industrial design, shoes, bags, and clothing items. GAN is also extensively used in the field
demonstrating
Electronics 2019, 8, xmemorization with SGD and a small learning rate [221].
FOR PEER REVIEW 38 of 67
of game development and artificial video generation [220]. GANs have two different areas of DL that
they fall into semi-supervised and unsupervised. Some research in these areas focuses on the
topology of the GAN architecture to improve functionality and the training approach. Deep
convolution GAN (DCGAN) is a convolution-based GAN approach proposed in 2015 [221]. This
semi-supervised approach has shown promised results compared to its unsupervised counterpart.
The regenerated results from DCGAN have shown in the following figures [183]. Figure 39 according
to article in [221], shows the output for generated bedroom images after one training pass through
the dataset. Most of the figures included in this section are generated through experiments.
Theoretically, the model could learn to memorize training examples, but this is experimentally
unlikely as we train with a small learning rate and mini batches with SGD. We are aware of no prior
empirical evidence demonstrating memorization with SGD and a small learning rate [221].
Figure 39. Experimental outputs of bedroom images.
Figure 40
Figure 40 represents
represents generated bedroom images
generated bedroom images after
after five
five epochs
epochs of
of training. There appears
training. There appears to
to
be evidence of visual under-fitting via repeated noise textures across multiple samples, such as
be evidence of visual under-fitting via repeated noise textures across multiple samples, such as the the
baseboards of
baseboards of some
some of
of the
the beds.
beds.
Figure
Figure 39.
39. Experimental
Experimental outputs
outputs of
of bedroom
bedroom images.
images.
Figure
Figure 4040 represents
represents generated
generated bedroom
bedroom images
images after
after five
five epochs
epochs of
of training.
training. There
There appears
appears toto
be
be evidence
evidence
Electronics
of
2019,of
visual
visual under-fitting via repeated noise textures across multiple samples, such as
8, 292
under-fitting via repeated noise textures across multiple samples, such the
as of
38 the
66
baseboards
baseboards of of some
some of
of the
the beds.
beds.
Figure
Figure40.
Figure 40.Reconstructed
40. Reconstructedbedroom
Reconstructed bedroomimages
bedroom images using
using deep
deep convolution
convolution GAN
GAN (DCGAN).
(DCGAN).
In
In Figure
Figure 40,
40, according
according toto article
article in
in [221],
[221], the
the top
top rows
rows interpolation
interpolation between
between aa series
between series ofof nine
nine
random
random points
pointsin
points in Z,
Z, and
and show
show that
that the
the learned
learned space
space has
has smooth
smooth transitions.
transitions. In In every
every image,
image, space
space
plausibly
plausibly looks
looks like
like aabedroom.
bedroom. InInthe
the6th
the 6throw,
6th row,you
row, yousee
seeaaaroom
see roomwithout
room withoutaaawindow
without windowslowly
window slowlytransforming
slowly transforming
transforming
into a room with a giant window.
into a room with a giant window. In In the
In the 10th
the 10th row,
row, you see what appears to be a TV slowly being
10th row, you see what appears to be a TV slowly being
transformed
transformed into
into
transformed into a aa window.
window. The
The following
following
following Figure
Figure 41
41 shows
shows the
the effective
effective application
application of
of latent
latent space
space
vectors. Latent space
Latent space
vectors. Latent vectors
space vectors can
vectors can be
can be turned
be turned into
turned into meaning
into meaning output
output by first performing addition and
meaning output by first performing addition and
subtraction
subtractionoperations
operationsfollowed
followed byby aa decode.
decode. Figure
Figure41 41according
according to to article
articleinin [221],
[221], shows
showsthat thataa man
man
with
with glasses
glasses minus
minus aa manman and
and add
add aa woman
woman whichwhich results
results in
results in aa woman
woman withwith glasses.
glasses.
Figure
Figure
Figure 41.Example
41.
41. Exampleof
Example ofsmile
of smile
smile arithmetic and
arithmetic and arithmeticfor
and arithmetic
arithmetic forwearing
for wearingglass
wearing glassusing
glass using
using GAN:
GAN:
GAN: a man
aa man
man with
with
with
glasses minus
glasses minusman without
man glasses
without plus
glasses woman
plus woman without
withoutglasses equal
glasses to
equal woman
to womanwith glasses.
with
glasses minus man without glasses plus woman without glasses equal to woman with glasses. glasses.
Figure 42, according

Figure according to article
article in [221],
[221], shows aa turn
turn vector was
was created from
from four averaged
averaged
Figure 42,42, according to to article in
in [221], shows
shows a turn vector
vector was created
created from fourfour averaged
samples
samples of faces looking left versus looking right. By adding interpolations along this axis of random
samples of of faces
faces looking
looking left
left versus
versus looking
looking right.
right. By
By adding
adding interpolations
interpolations along
along this
this axis
axis of
of random
random
samples
samples the pose can be reliably transformed. There are some interesting applications that have
samples thethe pose
pose can
can be
be reliably
reliably transformed.
transformed. There
There are
are some
some interesting
interesting applications
applications that
that have
have been
been
been proposedGANs.
proposed for GANs. example,
For example, natural indoor scenes are generatedwith
with improved GAN GAN
proposed for for GANs. For For example, natural
natural indoor
indoor scenes
scenes are
are generated
generated with improved
improved GAN
structures. TheseGANs
structures. GANs learnsurface
surface normaland and arecombined
combined withaaStyle
Style GANby by Wang and and
structures. These
These GANs learn learn surface normal
normal and areare combined with
with a Style GAN
GAN by Wang Wang and
Gupta [222]. In this implementation, authors considered the style and structure of GAN named
(S2 -GAN), which generates a surface normal map. This is an improved version of GAN. In 2016, an
information-theoretic extension to the GAN called InfoGAN was proposed. An infoGAN can learn
with better representations in a completely unsupervised manner. The experimental results show
that the unsupervised InfoGAN is competitive with representation learning with the fully supervised
learning approach [223].
In 2016, another new architecture was proposed by Im et al. [224] where the recurrent concept
is included with the adversarial network during training. Chen et. al. [225] proposed Info GAN
(iGAN) which allowed image manipulation interactively on a natural image manifold. Image to image
translation with conditional adversarial networks is proposed in 2017. Another improved version
of GANs named Coupled Generative Adversarial Network (CoGAN) is a learned joint distribution
of multi-domain images. The existing approach does not need tuples of corresponding images in
different domains in the training set [226]. Bidirectional Generative Adversarial Networks (BiGANs
are learned with inverse feature mapping and shown that the resulting learned feature representation
is included with the adversarial network during training. Chen et. al. [225] proposed Info GAN
(iGAN) which allowed image manipulation interactively on a natural image manifold. Image to
image translation with conditional adversarial networks is proposed in 2017. Another improved
version of GANs named Coupled Generative Adversarial Network (CoGAN) is a learned joint
distribution of multi-domain images. The existing approach does not need tuples of corresponding
Electronics 2019, 8, 292 39 of 66
images in different domains in the training set [226]. Bidirectional Generative Adversarial Networks
(BiGANs are learned with inverse feature mapping and shown that the resulting learned feature
representation
is useful for auxiliaryis useful for auxiliary
supervised supervised
discrimination discrimination
tasks, competitive tasks, competitive
with contemporary with
approaches to
contemporary
un-supervisedapproaches to un-supervised
and self-supervised and self-supervised
feature learning [227]. feature learning [227].
Figure 42.
Figure Facegeneration
42. Face generation in
in different
different angle
angle using
using GAN.
GAN.
Recently, Google
Recently, Google proposed
proposed extended
extended versions
versions of of GANs
GANs called
called Boundary
Boundary Equilibrium
Equilibrium Generative
Generative
Adversarial Networks (BEGAN) with a simple but robust architecture [228]. BEGAN has aa better
Adversarial Networks (BEGAN) with a simple but robust architecture [228]. BEGAN has better
training procedure
training procedure with
with fast
fast and
and stable
stable convergence.
convergence. The The concept
concept of of equilibrium
equilibrium helps
helps toto balance
balance the
the
power of the discriminator against the generator. In addition, it can balance
power of the discriminator against the generator. In addition, it can balance the trade-off between the trade-off between
image diversity
image diversity and
and visual
visual quality
quality [228].
[228]. Another
Anothersimilar
similarwork
workisiscalled
calledWasserstein
Wasserstein GANGAN (WGAN)
(WGAN)
algorithm that shows significant benefits over traditional GAN [229]. WGANs
algorithm that shows significant benefits over traditional GAN [229]. WGANs had two major benefits had two major
benefits over traditional GANs. First, a WGAN meaningfully correlates
over traditional GANs. First, a WGAN meaningfully correlates the loss metric with the generator’sthe loss metric with the
generator’s convergence
convergence and sample
and sample quality. quality.
Secondly, Secondly,
WGANs haveWGANs
improved have improved
stability of thestability of the
optimization
optimization
process. process.
The improved
The improved version
version of
ofWGAN
WGANisisproposed
proposedwith witha new
a new clipping
clippingtechnique,
technique,which
which penalizes
penalizes the
normal
the of the
normal ofgradient of the of
the gradient critic
thewith respect
critic with to its inputs
respect [230].
to its There
inputs is a promising
[230]. There is aarchitecture
promising
that has been proposed based on generative models where the images
architecture that has been proposed based on generative models where the images are are represented with untrained
represented
DNNuntrained
with that give anDNNopportunity
that give for
an better understanding
opportunity for betterand visualization and
understanding of DNNs [231]. Adversarial
visualization of DNNs
examples for generative models have also been introduced [232].
[231]. Adversarial examples for generative models have also been introduced [232].Energy-based GAN was proposed by
Energy-based
Yann LeCun
GAN from Facebook
was proposed by Yann inLeCun
2016 [233].
fromThe traininginprocess
Facebook is difficult
2016 [233]. for GANs,
The training Manifold
process Matching
is difficult for
GAN (MMGAN) proposed with better training process which
GANs, Manifold Matching GAN (MMGAN) proposed with better training process which is is experimented on three different
datasets and the
experimented on experimental
three different results clearly
datasets anddemonstrate
the experimentalthe efficacy
resultsofclearly
MMGAN against other
demonstrate the
models [234]. GAN for geo-statistical simulation and inversion with efficient
efficacy of MMGAN against other models [234]. GAN for geo-statistical simulation and inversion training approach [235].
with Probabilistic GANapproach
efficient training (PGAN) [235].
which is a new kind of GAN with a modified objective function. The
mainProbabilistic
idea behindGAN this method is
(PGAN) whichto integrate
is a newa probabilistic
kind of GANmodel with a(A Gaussian
modified Mixture
objective Model) The
function. into
the GAN
main ideaframework
behind thisthat supports
method likelihooda rather
is to integrate than classification
probabilistic [236]. AMixture
model (A Gaussian GAN with Bayesian
Model) into
Network model [237]. Variational Auto encode is a popular deep learning approach,
the GAN framework that supports likelihood rather than classification [236]. A GAN with Bayesian which is trained
with Adversarial Variational Bayes (AVB) which helps to establish a principle connection between VAE
and GAN [238]. The f-GAN which is proposed based on the general feed-forward neural network [239].
Markov model-based GAN for texture synthesis [240]. Another generative model based on the doubly
stochastic MCMC method [241]. GAN with multi-Generator [242]
Is an unsupervised GAN capable of learning on a pixel level domain adaptation that transforms
in the pixel space from one domain to another domain? This approach provides state-of-the-art
performance against several unsupervised domain adaptation techniques with a large margin [243].
A new network is proposed called Schema Network, which is an object-oriented generative physics
simulator able to disentangle multiple causes of events reasoning through causes to achieve a goal that
is learned from dynamics of an environment from data [244]. There is interesting research that has
been conducted with a GAN that is to Generate Adversarial Text to Image Synthesis. In this paper,
the new deep architecture is proposed for GAN formulation which can take the text description of
an image and produce realistic images with respect to the inputs. This is an effective technique for
Electronics 2019, 8, 292 40 of 66
text-based image synthesis using a character level text encoder and class conditional GAN. GAN is
evaluated on bird and flower dataset first then general text to the image which is evaluated on MS
COCO dataset [40].
7.2. Applications of GAN

This learning algorithm has been applied in the different domain of applications that are discussed
in the following sections:
7.2.1. GAN for Image Processing

GANs used for generating a photo-realistic image using a super-resolution approach [245]. GAN
for semantic segmentation with semi and weakly supervised approach [246]. Text Conditioned
Auxiliary Classifier GAN (TAC-GAN) which is used for generating or synthesizing images from a
text description [247]. Multi-style Generative network (MSG-Net) which retains the functionality of
optimization-based approaches with fast speed. This network matches image styles at multiple scales
and puts the computational burden into training [248]. Most of the time, vision systems struggle with
rain, snow, and fog. A single image de-raining system is proposed using a GAN recently [249].
7.2.2. GAN for Speech and Audio Processing

An End-to-End Dialogue system using Generative Hierarchical Neural Network models [250].
In addition, GANs have been used in the field of speech analysis. Recently, GANs are used for speech
enhancement which is called SEGAN that incorporates further speech-centric design to improve
performance progressively [251]. GAN for symbolic-domain and music generation which performs
comparably against Melody RNN [252].
7.2.3. GAN for Medical Information Processing

GANs for Medical Imagining and medical information processing [136], GANs for medical
image de-noising with Wasserstein distance and perceptual loss [253]. GANs can also be used for
segmentation of Brain Tumors with conditional GANs (cGAN) [254]. A General medical image
segmentation approach is proposed using a GAN called SegAN [255]. Before the deep learning
revolution, compressive sensing is one of the hottest topics. However, Deep GAN is used for
compressed sensing that automates MRI [256]. In addition, GANs can also be used in health record
processing, due to the privacy issue the electronic health record (EHR) is limited to or is not publicly
available like other datasets. GANs are applied to synthetic EHR data which could mitigate risk [257].
Time series data generation with Recurrent GAN (RGAN) and Recurrent Conditional GAN (RCGAN)
has been introduced [258]. LOGAN consists of the combination of a generative and discriminative
model for detecting the overfitting and recognition inputs. This technique has been compared against
state-of-the-art GAN technique, including GAN, DCGAN, BEGAN and a combination of DCGAN
with a VAE [259].
7.2.4. Other Applications

A new approach called Bayesian Conditional GAN (BC-GAN) which can generate samples from
deterministic inputs. This is simply a GAN with a Bayesian framework that can handle supervised,
semi-supervised and unsupervised learning problems [260,261]. In machine learning and deep learning
community, online learning is an important approach. GANs are used for online learning in which it is
being trained for finding a mixed strategy in a zero-sum game which is named Checkov GAN 1 [262].
Generative moment matching networks based on statistical hypothesis testing called maximum mean
discrepancy (MMD) [263]. One of the interesting ideas to replace the discriminator of GAN with
two-sample based kernel MMD is called MMD-GAN. This approach significantly outperforms the
Electronics 2019, 8, 292 41 of 66
Generative moment matching network (GMMN) technique which is an alternative approach for the
generative model [264].
Some other applications of GAN include pose estimation [265], photo editing network [266], and
anomaly detection [267]. DiscoGAN for learning cross-domain relation with GAN [40], unsupervised
image-to-image translation with generative model, [268], single shot learning with GAN [269], response
generation and question answering system [270,271]. Last but not least, WaveNet as a generative
model has been developed for generating audio waveform in [272] and dual path network in [273].
8. Deep Reinforcement Learning (DRL)

In the previous sections, we have focused on supervised and unsupervised deep learning
approaches, including DNN, CNN, RNN including LSTM and GRU, AE, RBM, GAN etc. These
types of deep learning approaches are used for prediction, classification, encoding, decoding, data
generation, and many more application domains. However, this section demonstrates a survey on
Deep Reinforcement Learning (DRL) based on the recently developed methods in this field of RL.
8.1. Review on DRL

DRL is a learning approach which learns to act with general sense from the unknown real
environment (For details please read the following article [46,274]). The conceptual diagram for DRL
approach is shown in Figure 43. RL can be applied in a different scope of field, including fundamental
Sciences for decision making, Machine learning from a computer science point of view, in the field of
engineering and mathematics, optimal control, robotics control, power station control, wind turbines,
and Neuroscience the reward strategy is widely studied in the last couple of decades. It is also applied
in economic utility or game theory for making better decisions and for investment choices. The
psychological concept of classical conditioning is how animals learn. Reinforcement learning is a
technique for what to do and how to match a situation to an action. Reinforcement learning is different
from supervised learning technique and other kinds of learning approaches studies recently, including
traditional machine learning, statistical pattern recognition, and ANN.
Figure 43.
Figure 43. Conceptual diagram for
for Reinforcement
Reinforcement Learning
Learning (RL)
(RL) system.
system.
Unlike
Unlike the
the general
general supervised
supervised andand unsupervised
unsupervised machine learning, RL is defined not by
characterizing
characterizing learning methods,
methods, but
but by
by characterizing
characterizing aa learning
learning problem.
problem. However, the recent recent
success
success of
ofDL
DLhas
hashad
hada ahuge
hugeimpact
impactononthe success
the of of
success DRLDRL which is known
which as DRL.
is known as DRL.According to the
According to
learning strategy,
the learning the RL the
strategy, technique is learnedisthrough
RL technique learnedobservation. For observing
through observation. Fortheobserving
environment,the
the promisingthe
environment, DLpromising
techniques DLinclude CNN,
techniques RNN,CNN,
include LSTM, and LSTM,
RNN, GRU are andused
GRU depending upon the
are used depending
observation space. As space.
upon the observation DL techniques encode data
As DL techniques efficiently,
encode therefore, therefore,
data efficiently, the following step of action
the following step
is
of performed more accurately.
action is performed According
more accurately. to the action,
According to thethe agentthe
action, receives an appropriate
agent receives reward
an appropriate
respectively. As a result,
reward respectively. As athe entire
result, theRL approach
entire becomes
RL approach more efficient
becomes to learntoand
more efficient interact
learn in the
and interact
environment with better
in the environment with performance.
better performance.
However, the history of the modern DRL revolution began from Google Deep Mind in 2013 with
Atari games with DRL. In which the DRL based approaches perform better against the human expert
in almost all of the games. In this case, the environment is observed on video frames which are
processed using a CNN [275,276]. The success of DRL approaches depends on the level of difficulty
of the task attempt to be solved. After a huge success of Alpha-Go and Atari from Google Deep mind,
they proposed a reinforcement learning environment based on StarCraft II in 2017, which is called
Electronics 2019, 8, 292 42 of 66
However, the history of the modern DRL revolution began from Google Deep Mind in 2013
with Atari games with DRL. In which the DRL based approaches perform better against the human
expert in almost all of the games. In this case, the environment is observed on video frames which are
processed using a CNN [275,276]. The success of DRL approaches depends on the level of difficulty of
the task attempt to be solved. After a huge success of Alpha-Go and Atari from Google Deep mind,
they proposed a reinforcement learning environment based on StarCraft II in 2017, which is called
SC2LE (StarCraft II Learning Environment) [277]. The SC2LE is a game with multi-agent with multiple
players’ interactions. This proposed approach has a large action space involving the selection and
control of hundreds of units. It contains many states to observe from raw feature space and it uses
strategies over thousands of steps. The open source Python-based StarCraft II game engine has been
provided free in online.
8.2. Q-Learning
There are some fundamental strategies which are essential to know for working with DRL. First,
the RL learning approach has a function that calculates the Quality of state-action combination which
is called Q-Learning (Q-function). Algorithm 2 describes the basic computational flow of Q-learning.
Q-learning is defined as a model-free reinforcement learning approach which is used to find
an optimal action-selection policy for any given (finite) Markov Decision Process (MDP). MDP is a
mathematical framework for modeling decision using state, action and rewards. Q-learning only needs
to know about the states available and what are the possible actions in each state. Another improved
version of Q-Learning known as Bi-directional Q-Learning. In this article, the Q-Learning is discussed,
for details on bi-directional Q-Learning please see Reference [278].
At each step s, choose the action which maximizes the following function Q (s, a)
• Q is an estimated utility function—it tells us how good an action is given in a certain state
• r (s, a) immediate reward for making an action best utility (Q) for the resulting state
This can be formulated with the recursive definition as follows:
Q(s, a) = r(s, a) + γ maxa0 Q s0 , a0 .

(59)
This equation is called Bellman’s equation, which is the core equation for RL. Here r(s, a) is the
immediate reward, γ is the relative value of delay vs. immediate rewards [0, 1] s0 is the new state after
action a. The a and a0 are an action in sate s and s0 respectively. The action is selected based on the
following equation:
π (s) = argmaxa Q(s, a). (60)
In each state, a value is assigned called a Q-value. When we visit a state and we receive a reward
accordingly. We use the reward to update the estimated value for that state. As the reward is stochastic,
as a result, we need to visit the states many times. In addition, it is not guaranteed that we will get
the same reward (Rt ) in another episode. The summation of the future rewards in episodic tasks
and environments are unpredictable, further in the future, we go further with the reward diversely
as expressed,
Gt = Rt+1 + Rt+2 + Rt+3 + . . . . . . . . . . . + RT . (61)
The sum of discounted future rewards in both cases are some factor as scalar.
Gt = γ Rt+1 + γ2 Rt+2 + γ3 Rt+3 + . . . . . . . . . . . + γT RT , (62)
here γ is a constant. The more we are in the future, the less we take the reward into account.
Properties of Q-learning:
• Convergence of Q-function: Approximation will be converged to the true Q-function, but it must
visit possible state-action pair infinitely many times.
Electronics 2019, 8, 292 43 of 66
• The state table size can be vary depending on the observation space and complexity.
• Unseen values are not considered during observation.
The way to fix these problems is to use a neural network (particularly DNN) as an approximation
instead of the state table. The inputs of DNN are the state and action and the outputs are numbers
between 0 and 1 that represent the utility encoding the states and actions properly. That is the
place where the deep learning approaches contribute to making better decisions with respect to
the state information. Most of the cases for observing the environment, we use several acquisition
devices, including a camera or other sensing devices for observing the learning environment. For
example, if you observed the setup for the challenge of Alpha-Go then it can be seen that the
environment, action, and reward are learned based on the pixel values (pixel in action). For details see
References [275,276,279].
However, it is difficult to develop an agent which can interact or perform well in any observation
environment. Therefore, most of the researchers in the field select their action space or environment
before training the agent for that environment. The benchmark concept, in this case, is a little bit
different compared to supervised or unsupervised deep learning approach. Due to the variety of
environments, the benchmark depends on what level of difficulty the environment has been considered
compared to the previous or exiting researches? The difficulties depend on the different parameters,
number of agents, a way of interaction between the agents, the number of players and so on.
Recently, another good learning approach has been proposed for DRL [46,274]. There are many
papers published with different networks of DRL, including Deep Q-Networks (DQN), Double DQN,
Asynchronous methods, policy optimization strategy (including deterministic policy gradient, deep
deterministic policy gradient, guided policy search, trust region policy optimization, combining policy
gradient and Q-learning) are proposed [46,274]. Policy Gradient (DAGGER) Superhuman GO using
supervised learning with policy gradient and Monte Carlo tree search with value function [46,280].
Robotics manipulation using guided policy search [281]. DRL for 3D games using policy gradients [282].
Algorithm 2: Q-Learning
Initialization:
For each state-action pair (s, a)
initialize the table entry Q̂(s, a) to zero
Steps:
1. Observed the current state s
2. REPEAT:
- Select an action a and execute it

- Received immediate reward r
- Observe the new state s0
- Update the table entry for Q̂(s, a) as follows:
Q̂(s, a) = r + γ max a0 ( Q(s0 , a0 ))
- s = s0
8.3. Recent Trends of DRL with Applications

There is a survey published recently, where basic RL, DRL DQN, trust region policy optimization,
and asynchronous advantage actor-critic are proposed. This paper also discusses the advantages of
deep learning and focuses on visual understanding via RL and the current trend of research [283]. A
network cohesion constrained based on online RL techniques is proposed for health care on mobile
devices called mHealth. This system helps similar users to share information efficiently to improve
and convert the limited user information into better-learned policies [284]. Similar work with the
Electronics 2019, 8, 292 44 of 66
group-driven RL is proposed for health care on a mobile device for personalized mHealth Intervention.
In this work, K-means clustering is applied for grouping the people and finally shared with RL
policy for each group [285]. Optimal policy learning is a challenging task with RL for an agent.
Option-Observation Initiation sets (OOIs) allow agents to learn optimal policies in the challenging
task of POMDPs which are learned faster than RNN [286]. 3D Bin Packing Problem (BPP) is proposed
with DRL. The main objective is to place the number of the cuboid-shaped items that can minimize the
surface area of the bin [287].
The import component of DRL is the reward which is determined based on the observation
and the action of the agent. The real-world reward function is not perfect at all times. Due to the
sensor error, the agent may get maximum reward whereas the actual reward should be smaller. This
paper proposed a formulation based on generalized Markov Decision Problem (MDP) called Corrupt
Reward MDP [288]. The trust region optimization based deep RL is proposed using recently developed
Kronecker-factored approximation to the curvature (K-FAC) [289]. In addition, there is some research
that has been conducted in the evaluation of physics experiments using the deep learning approach.
This experiment focuses agent to learn basic properties, such as mass and cohesion of the objects in the
interactive simulation environment [290].
Recently Fuzzy RL policies have been proposed that is suitable for continuous state and action
space [291]. The important investigation and discussion are made for hyper-parameters in policy
gradient for continuous control, the general variance of the algorithm. This paper also provides a
guideline for reporting results and comparison against baseline methods [292]. Deep RL is also applied
to high precision assembly tasks [293]. The Bellman equation is one of the main functions of RL
technique, a function approximation is proposed which ensures that the Bellman Optimality Equation
always holds. Then the function is estimated to maximize the likelihood of the observed motion [294].
DRL based hierarchical system is used for could resource allocation and power management in could
computing system [295]. A novel Attention-aware Face Hallucination (Attention-FC) is proposed
where Deep RL is used for enhancing the quality of the image on a single patch for images which are
applied to face images [296].
9. Bayesian Deep Learning (BDL)

The DL approaches have been providing state-of-the-art accuracy for different applications. However,
DL approaches are unable to deal with the uncertainty of a given task due to model uncertainty. These
learning approaches take input and assume the class probability without justification [297,298]. In
2015, two African American humans recognized as gorilla with an image classification system [299].
There are several application domains where the uncertainty can be raised, including self-driven car,
bio-medical applications. However, the BDN, which is an intersection between DL and Bayesian
probability approaches show better results in different applications and understand the uncertainty
of problems, including multi-task problems [297,298]. The uncertainty is estimated with applying
probability distribution over the model weights or mapping the outputs’ probability [297,298].
The BDL is becoming very popular among the DL research community. In addition, the BDL
approaches have been proposed with CNN techniques where the probability distribution is applied to
weight. These techniques help to deal with model overfitting problem and lack of training samples
which are the two commons challenges for DL approaches [300,301]. Finally, there are some other
research papers have published recently where some advanced techniques have been proposed on
BDL [302–305].
10. Transfer Learning
10.1. Transfer Learning

A good way to explain transfer learning is to look at the student-teacher relationship. A teacher
offers a course after gathering details knowledge regarding that subject [48]. The information will be
justification [297,298]. In 2015, two African American humans recognized as gorilla with an image
classification system [299]. There are several application domains where the uncertainty can be
raised, including self-driven car, bio-medical applications. However, the BDN, which is an
intersection between DL and Bayesian probability approaches show better results in different
applications
Electronics 2019, 8,and
292 understand the uncertainty of problems, including multi-task problems [297,298].
45 of 66
The uncertainty is estimated with applying probability distribution over the model weights or
mapping the outputs’ probability [297,298].
conveyed through a series of lectures over time. This can be considered that the teacher (expert) is
The BDL is becoming very popular among the DL research community. In addition, the BDL
transferring information (knowledge) to the students (learner). The same thing happens in case of
approaches have been proposed with CNN techniques where the probability distribution is applied
deep learning, a network is trained with a big amount data and during the training, the model learns
to weight. These techniques help to deal with model overfitting problem and lack of training samples
the weights and bias. These weights can be transferred to other networks for testing or retraining a
which are the two commons challenges for DL approaches [300,301]. Finally, there are some other
similar new model. The network can start with pre-trained weights instead of training from scratch.
research papers have published recently where some advanced techniques have been proposed on
The conceptual diagram for transfer learning method is shown in Figure 44.
BDL [302–305].
Figure 44. Conceptual diagram for transfer learning: Pretrained on ImageNet and transfer learning is
Figure 44. Conceptual diagram for transfer learning: Pretrained on ImageNet and transfer learning is
used for retraining on PASCAL dataset.
used for retraining on PASCAL dataset.
10.2. What Is A Pre-trained Model?
10. Transfer Learning
A pre-trained model is a model which is already trained in the same domains as the intended
domain. For example,
10.1. Transfer Learningfor an image recognition task, an Inception model already trained on ImageNet
can be downloaded. The Inception model can then be used for a different recognition task, and instead
A good
of training way to
it from explain
scratch thetransfer
weightslearning
can be is to as
left look at thesome
is with student-teacher relationship.
learned features. A teacher
This method of
offers a course after gathering details knowledge regarding that subject [48]. The information
training is useful when there is a lack of sample data. There are a lot of pre-trained models available will be
conveyed VGG,
(including through a series
ResNet, of Inception
and lectures over
Net time. This can
on different be considered
datasets) that thefrom
in model-zoo teacher (expert) is
the following
transferring
link: information (knowledge) to the students (learner). The same thing happens in case of
https://github.com/BVLC/caffe/wiki/Model-Zoo.
deep learning, a network is trained with a big amount data and during the training, the model learns
the weights
10.3. Why WillandYoubias.
Use These weights
Pre-trained can be transferred to other networks for testing or retraining a
Models?
similar new model. The network can start with pre-trained weights instead of training from scratch.
There are a lot of reasons for using pre-trained models. Firstly, it requires a lot of expensive
The conceptual diagram for transfer learning method is shown in Figure 44.
computation power to train big models on big datasets. Secondly, it can take up to multiple weeks to
train big models. Training new models with pre-trained weights can speed up convergence, as well as
help the network generalization.
10.4. How Will You Use Pre-trained Models?

We need to consider the following criterions with respective application domains and size of the
dataset when using the pre-trained weights which are shown in Table 3.
Table 3. Criterions need to be considered for transfer learning.
Methods New Dataset but Small New Dataset but Large

Pre-trained model on similar but new Freeze weights and train linear classifier Fine-tune all the layers (pre-train for faster
dataset from top level features convergence and better generalization)
Pre-trained model a on different but new Freeze weights and train linear classifier Fine-tune all the layers (pre-train for
dataset from non-top-level features enhanced convergence speed)
Electronics 2019, 8, 292 46 of 66
10.5. Working with Inference

Research groups working specifically on inference applications look into optimization approaches
that include model compression. Model compression is important in the realm of mobile devices or
special purpose hardware because it makes models more energy efficient, as well as faster.
10.6. The Myth about Deep Learning

There is a myth; do you need a million labeled samples for training a deep learning model?
The answer is yes but, in most cases, the transfer learning approach is used to train deep learning
approaches without having large amounts of label data. For example, the following Figure 44
demonstrates the strategy for the transfer learning approach in details. Here the primary model
has been trained with a large amount of labeled data which is ImageNet and then the weights are used
to train with the PASCAL dataset. The actual reality is:
• Possible to learn useful representations from unlabeled data.

• Transfer learning can help learned representation from the related task [306].
We can take a trained network for a different domain which can be adapted for any other domain
for the target task [307,308]. First training a network with a close domain for which it is easy to get
labeled data using standard backpropagation, for example, ImageNet classification, pseudo classes
from augmented data. Then cut off the top layers of network and replace with the supervised objective
for the target domain. Finally, tune the network using backpropagation with labels for the target
domain until validation loss starts to increase [307,308]. There are some survey papers and books that
are published on transfer learning [309,310]. Self-taught learning with transfer learning [311]. Boosting
approach for transfer learning [312].
11. Energy Efficient Approaches and Hardware for DL
11.1. Overview
DNNs have been successfully applied and achieved better recognition accuracies in different
application domains, such as computer vision, speech processing, natural language processing, big
data problem and many more. However, most of the cases the training is being executed on Graphics
Processing Units (GPU) for dealing with big volumes of data which is expensive in terms of power.
Recently researchers have been training and testing with deeper and wider networks to achieve
even better classification accuracy to achieve human or beyond human level recognition accuracy
in some cases. While the size of the neural network is increasing, it becomes more powerful and
provides better classification accuracy. However, the storage consumption, memory bandwidth
and computational cost are increasing exponentially. On the other hand, these types of massive
scale implementation with large numbers of network parameters are not suitable for low power
implementation, unmanned aerial vehicle (UAV), different medical devices, a low memory system,
such as mobile devices, Field Programmable Gate Array (FPGA) and so on.
There is much research going on to develop better network structures or networks with lower
computation cost, fewer numbers of parameters for low-power and low-memory systems without
lowering classification accuracy. There are two ways to design an efficient deep network structure:
• The first approach is to optimize the internal operational cost with an efficient network structure;
• Second design a network with low precision operations or a hardware efficient network.
The internal operations and parameters of a network structure can be reduced by using low
dimensional convolution filters for convolution layers [71,99].
There is a lot of benefit to this approach. Firstly, the convolutional with rectification operations
makes the decision more discriminative. Secondly, the main benefit of this approach is to reduce
the number of computation parameters drastically. For example, if one layer has 5 × 5 dimensional
Electronics 2019, 8, 292 47 of 66
filters which can be replaced with two 3 × 3 dimensional filters (without pooling layer in between
then) for better feature learning; three 3 × 3 dimensional filters can be used as a replacement of 7 × 7
dimensional filters and so on. Benefits of using a lower-dimensional filter are that assuming both
the present convolutional layer has C channels, for three layers for 3 × 3 filter the total number of
parameters are weights: 3 × (3 × 3 × C × C) = 27C2 weights, whereas in the size of the filter is 7 × 7,
the total number of parameters are (7 × 7 × C × C) = 49C2 , which is almost double compared to the
three 3 × 3 filter parameters. Moreover, placement of layers, such as convolutional, pooling, drop-out
in the network in different intervals has an impact on overall classification accuracy. There are some
strategies that are mentioned to optimize the network architecture recently to design robust deep
learning models [99,100,313] and efficient implementation of CNNs on FPGA platform [314].
Strategy 1: Replace the 3 × 3 filter with 1 × 1 filters. The main reasons to use a lower dimension
filter to reduce the overall number of parameter. By replacing 3 × 3 filters with 1 × 1 can be reduced
9x number of parameters.
Strategy 2: Decrease the number of input channels to 3 × 3 filters. For a layer, the sizes of the
output feature maps are calculated, which is related to the network parameters using N − F
S + 1, where
N is input map’s size, F is filter size, S is for strides. To reduce the number of parameters, it is not only
enough to reduce the size of the filters, but also it requires controlling number of input channels or
featuring dimension.
Strategy 3: Down-sample late in the network so that convolution layers have activation maps:
The outputs of present convolution layers can be at least 1 × 1 or often larger than 1 × 1. The output
width and height can be controlled by some criterions: (1) The size of the input sample (e.g., 256 × 256)
and (2) Choosing the post down sample layer. Most commonly pooling layers are such as average or
max pooling layer is used, there is an alternative sub-sampling layer with convolution (3 × 3 filters)
and stride with 2. If most of the earlier layers have larger stride, then most of the layers will have small
numbers of activation maps.
11.2. Binary or Ternary Connect Neural Networks

The computation cost can be reduced drastically with the low precision of multiplication and
few multiplications with drop connection [315,316]. These papers also introduced on Binary Connect
Neural Networks (BNN) Ternary Connect Neural Networks (TNN). Generally, multiplication of a
real-valued weight by a real-valued activation (in the forward propagations) and gradient calculation
(in the backward propagations) are the main operations of deep neural networks. Binary connect or
BNN is a technique that eliminates the multiplication operations by converting the weights used in
the forward propagation to be binary, i.e., constrained to only two values (0 and 1 or -1 and 1). As
a result, the multiplication operations can be performed by simple additions (and subtractions) and
makes the training process faster. There are two ways to convert real values to its corresponding
binary values, such as deterministic and stochastic. In the case of the deterministic technique,
straightforward thresholding technique is applied to weights. An alternative way to do that is a
stochastic approach where a matrix is converted to binary based on probability where the hard sigmoid
function is used because it is computationally inexpensive. The experimental result shows significantly
good recognition accuracy [317–319]. There are several advantages of BNN as follows:
• It is observed that the binary multiplication on GPU is almost seven times faster than traditional
matrix multiplication on GPU
• In forward pass, BNNs drastically reduce memory size and accesses, and replace most arithmetic
operation with bit-wise operations, which lead great increase of power efficiency
• Binarized kernels can be used in CNNs which can reduce around 60% complexity of
dedicated hardware.
• It is also observed that memory accesses typically consume more energy compared to the
arithmetic operation and memory access cost increases with memory size. BNNs are beneficial
with respect to both aspects.
Electronics 2019, 8, 292 48 of 66
There are some other techniques that have been proposed in the last few years [320–323]. Another
power efficient and hardware friendly network structure has been proposed for a CNN with XNOR
operations. In XNOR based CNN implementations, both the filters and input to the convolution layer
is binary. This result about 58x faster convolutional operations and 32x memory saving. In the same
paper, Binary-Weight-Networks was proposed which saved around 32x memory saving. That makes
it possible to implement state-of-the-art networks on CPU for real-time use instead of GPU. These
networks are tested on the ImageNet dataset and provide only 2.9% less classification accuracy than
full-precision AlexNet (in top-1% measure). This network requires less power and computation time.
This could make it possible to accelerate the training process of deep neural network dramatically
for specialized hardware implementation [273,274]. For the first time, Energy Efficient Deep Neural
Network (EEDN) architecture was proposed for the neuromorphic system in 2016. In addition, they
released a deep learning framework called EEDN, which provides close accuracy to state-of-the-art
accuracy almost all the popular benchmarks except ImageNet dataset [324,325].
12. Hardware for DL

Along with the algorithmic development of DL approaches, there are many hardware architectures
have been proposed in the past few years [326]. The details about present trends of hardware for
deep learning have been published recently [49,326]. MIT proposed Eyeriss as hardware for deep
convolutional neural networks (DCNN) [327]. There is another architecture for machine learning called
Dadiannao [328]. Google developed hardware named Tensor Processing Unit (TPU) for deep learning
and was released in 2017 [329]. In 2016, efficient hardware that works for inference was released
and proposed by Stanford University called Efficient Inference Engine (EIE) [330]. IBM released a
neuromorphic system called TrueNorth in 2015 [324].
Deep learning approaches are not limited to the HPC platform, there is a lot of application already
developed which run on mobile devices. Mobile platforms provide data that is relevant to everyday
activities of the user, which can make a mobile system more efficient and robust by retraining the
system with collected data. There is some research ongoing to develop hardware friendly algorithms
for DL [331–333].
13. Other topics

There are several important topics, including frameworks, SDK, benchmark datasets, related
Journals and Conferences are included in Appendix A.
14. Summary
In this paper, we have provided an in-depth review of deep learning and its applications over
the past few years. Different state-of-the-art deep learning models in different categories of learning,
including supervised, unsupervised, and Reinforcement Learning (RL), as well as their applications in
different domains were reviewed. In addition, we have explained in detail the different supervised deep
learning techniques, including DNN, CNN, and RNN. The un-supervised deep learning techniques,
including AE, RBM, and GAN, were reviewed in detail. In the same section, we have considered
and explained unsupervised learning techniques which are proposed based on LSTM and RL. In
Section 8, we presented a survey on Deep Reinforcement Learning (DRL) with the fundamental
learning technique called Q-Learning. The recently developed Bayesian Deep Learning (BDL) and
Transfer Learning (TL) approaches are also discussed in Sections 9 and 10, respectively. Furthermore,
we have conducted a survey on energy efficient deep learning approaches, transfer learning with DL,
and hardware development trends of DL. Moreover, we have discussed some DL frameworks and
benchmark datasets, which are often used for the implementation and evaluation of deep learning
approaches. Finally, we have included relevant journals and conferences, where the DL community
has been publishing their valuable research articles.
Electronics 2019, 8, 292 49 of 66
Funding: This work was supported by the National Science Foundation under awards 1718633 and 1309708.
Acknowledgments: We would like to thank all authors mentioned in the reference of this paper from whom we
have learned a lot and thus made this review paper possible.
Conflicts of Interest: The authors declare no conflict of interest.
Appendix A
Most of the time people use different deep learning frameworks and Standard Development Kits
(SDKs) for implementing deep learning approaches which are listed below:
A.1. Frameworks
• Tensorflow: https://www.tensorflow.org/
• Caffe: http://caffe.berkeleyvision.org/
• KERAS: https://keras.io/
• Theano: http://deeplearning.net/software/theano/
• Torch: http://torch.ch/
• PyTorch: http://pytorch.org/
• Lasagne: https://lasagne.readthedocs.io/en/latest/
• DL4J (DeepLearning4J): https://deeplearning4j.org/
• Chainer: http://chainer.org/
• DIGITS: https://developer.nvidia.com/digits
• CNTK (Microsoft): https://github.com/Microsoft/CNTK
• MatConvNet: http://www.vlfeat.org/matconvnet/
• MINERVA: https://github.com/dmlc/minerva
• MXNET: https://github.com/dmlc/mxnet
• OpenDeep: http://www.opendeep.org/
• PuRine: https://github.com/purine/purine2
• PyLerarn2: http://deeplearning.net/software/pylearn2/
• TensorLayer: https://github.com/zsdonghao/tensorlayer
• LBANN: https://github.com/LLNL/lbann
A.2. SDKs
• cuDNN: https://developer.nvidia.com/cudnn
• TensorRT: https://developer.nvidia.com/tensorrt
• DeepStreamSDK: https://developer.nvidia.com/deepstream-sdk
• cuBLAS: https://developer.nvidia.com/cublas
• cuSPARSE: http://docs.nvidia.com/cuda/cusparse/
• NCCL: https://devblogs.nvidia.com/parallelforall/fast-multi-gpu-collectives-nccl/
A.3. Benchmark Datasets

Here is the list of benchmark datasets that are used often to evaluate deep learning approaches in
different domains of application:
A.3.1. Image Classification or Detection or Segmentation

List of datasets are used in the field of image processing and computer vision:
• MNIST: http://yann.lecun.com/exdb/mnist/
• CIFAR 10/100: https://www.cs.toronto.edu/~{}kriz/cifar.html
• SVHN/ SVHN2: http://ufldl.stanford.edu/housenumbers/
Electronics 2019, 8, 292 50 of 66
• CalTech 101/256: http://www.vision.caltech.edu/Image_Datasets/Caltech101/

• STL-10: https://cs.stanford.edu/~{}acoates/stl10/
• NORB: http://www.cs.nyu.edu/~{}ylclab/data/norb-v1.0/
• SUN-dataset: http://groups.csail.mit.edu/vision/SUN/
• ImageNet: http://www.image-net.org/
• National Data Science Bowl Competition: http://www.datasciencebowl.com/
• COIL 20/100: http://www.cs.columbia.edu/CAVE/software/softlib/coil-20.php
• MS COCO DATASET: http://mscoco.org/
• MIT-67 scene dataset: http://web.mit.edu/torralba/www/indoor.html
• Caltech-UCSD Birds-200 dataset: http://www.vision.caltech.edu/visipedia/CUB-200-2011.html
• Pascal VOC 2007 dataset: http://host.robots.ox.ac.uk/pascal/VOC/voc2007/
• H3D Human Attributes dataset: https://www2.eecs.berkeley.edu/Research/Projects/CS/vision/
shape/poselets/
• Face recognition dataset: http://vis-www.cs.umass.edu/lfw/
• For more data-set visit: https://www.kaggle.com/
• http://homepages.inf.ed.ac.uk/rbf/CVonline/Imagedbase.htm
• Recently Introduced Datasets in Sept. 2016:
• Google Open Images (~9M images)—https://github.com/openimages/dataset
• Youtube-8M (8M videos: https://research.google.com/youtube8m/
A.3.2. Text Classification

• Reuters-21578 Text Categorization Collection: http://kdd.ics.uci.edu/databases/reuters21578/
reuters21578.html
• Sentiment analysis from Stanford: http://ai.stanford.edu/~{}amaas/data/sentiment/
• Movie sentiment analysis from Cornel: http://www.cs.cornell.edu/people/pabo/movie-review-data/
A.3.3. Language Modeling

• Free eBooks: https://www.gutenberg.org/
• Brown and stanford corpus on present americal english: https://en.wikipedia.org/wiki/Brown_
Corpus
• Google 1Billion word corpus: https://github.com/ciprian-chelba/1-billion-word-language-
modeling-benchmark
A.3.4. Image Captioning

Flickr-8k: http://nlp.cs.illinois.edu/HockenmaierGroup/8k-pictures.html
• Flickr-30k
• Common Objects in Context (COCO): http://cocodataset.org/#overview, http://sidgan.me/
technical/2016/01/09/Exploring-Datasets
A.3.5. Machine Translation

• Pairs of sentences in English and French: https://www.isi.edu/natural-language/download/
hansard/
• European Parliament Proceedings parallel Corpus 196-2011: http://www.statmt.org/europarl/
• The statistics for machine translation: http://www.statmt.org/
A.3.6. Question Answering

• Stanford Question Answering Dataset (SQuAD): https://rajpurkar.github.io/SQuAD-explorer/
Electronics 2019, 8, 292 51 of 66
• Dataset from DeepMind: https://github.com/deepmind/rc-data

• Amazon dataset: http://jmcauley.ucsd.edu/data/amazon/qa/, http://trec.nist.gov/data/
qamain..., http://www.ark.cs.cmu.edu/QA-data/, http://webscope.sandbox.yahoo.co...,
http://blog.stackoverflow.com/20..
A.3.7. Speech Recognition

• TIMIT: https://catalog.ldc.upenn.edu/LDC93S1
• Voxforge: http://voxforge.org/
• Open Speech and Language Resources: http://www.openslr.org/12/
A.3.8. Document Summarization

• https://archive.ics.uci.edu/ml/datasets/Legal+Case+Reports
• http://www-nlpir.nist.gov/related_projects/tipster_summac/cmp_lg.html
• https://catalog.ldc.upenn.edu/LDC2002T31
A.3.9. Sentiment Analysis:

• IMDB dataset: http://www.imdb.com/
A.3.10. Hyperspectral Image Analysis

• http://www.ehu.eus/ccwintco/index.php/Hyperspectral_Remote_Sensing_Scenes
• https://engineering.purdue.edu/~{}biehl/MultiSpec/hyperspectral.html
• http://www2.isprs.org/commissions/comm3/wg4/HyRANK.html
In addition, there is another alternative solution in data programming that labels subsets of data
using weak supervision strategies or domain heuristics as labeling functions even if they are noisy and
may conflict samples [87].
A.4. Journals and Conferences

In general, researchers publish their primary version of research on the ArXiv (https://arxiv.org/).
Most of the conferences have been accepting papers on Deep learning and its related field. Popular
conferences are listed below:
A.4.1. Conferences
• Neural Information Processing System (NIPS)
• International Conference on Learning Representation (ICLR): What are you doing for
Deep Learning?
• International Conference on Machine Learning (ICML)
• Computer Vision and Pattern Recognition (CVPR): What are you doing with Deep Learning?
• International Conference on Computer Vision (ICCV)
• European Conference on Computer Vision (ECCV)
• British Machine Vision Conference (BMVC)
A.4.2. Journal
• Journal of Machine Learning Research (JMLR)
• IEEE Transaction of Neural Network and Learning System (ITNNLS)
• IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)
• Computer Vision and Image Understanding (CVIU)
• Pattern Recognition Letter
Electronics 2019, 8, 292 52 of 66
• Neural Computing and Application

• International Journal of Computer Vision
• IEEE Transactions on Image Processing
• IEEE Computational Intelligence Magazine
• Proceedings of IEEE
• IEEE Signal Processing Magazine
• Neural Processing Letter
• Pattern Recognition
• Neural Networks
• ISPPRS Journal of Photogrammetry and Remote Sensing
A.4.3. Tutorials on Deep Learning

• http://deeplearning.net/tutorial/
• http://deeplearning.stanford.edu/tutorial/
• http://deeplearning.net/tutorial/deeplearning.pdf
• Courses on Reinforcement Learning: http://rll.berkeley.edu/deeprlcourse/
A.4.4. Books on Deep Learning

• https://github.com/HFTrader/DeepLearningBook
• https://github.com/janishar/mit-deep-learning-book-pdf
• http://www.deeplearningbook.org/
References
1. Schmidhuber, J. Deep Learning in Neural Networks: An Overview. Neural Netw. 2015, 61, 85–117. [CrossRef]
[PubMed]
2. Bengio, Y.; LeCun, Y.; Hinton, G. Deep Learning. Nature 2015, 521, 436–444.
3. Bengio, Y.; Courville, A.; Vincent, P. Representation Learning: A Review and New Perspectives. IEEE Trans.
Pattern Anal. Mach. Intell. 2013, 35, 1798–1828. [CrossRef] [PubMed]
4. Bengio, Y. Learning deep architectures for AI. Found. Trends Mach. Learn. 2009, 2, 1–127. [CrossRef]
5. Mnih, V.; Kavukcuoglu, K.; Silver, D.; Rusu, A.A.; Veness, J.; Bellemare, M.G.; Graves, A.; Riedmiller, M.;
Fidjeland, A.K.; Ostrovski, G.; et al. Human-level control through deep reinforcement learning. Nature 2015,
518, 529–533. [CrossRef] [PubMed]
6. Mnih, V.; Kavukcuoglu, K.; Silver, D.; Graves, A.; Antonoglou, I.; Wierstra, D.; Riedmiller, M. Playing atari
with deep reinforcement learning. arXiv 2013, arXiv:1312.5602.
7. Krizhevsky, A.; Sutskever, I.; Hinton, G.E. Imagenet classification with deep convolutional neural networks.
In Proceedings of the 25th International Conference on Neural Information Processing Systems, Lake Tahoe,
NV, USA, 3–6 December 2012; pp. 1106–1114.
8. Zeiler, M.D.; Fergus, R. Visualizing and understanding convolutional networks. arXiv 2013, arXiv:1311.2901.
9. Simonyan, K.; Zisserman, A. deep convolutional networks for large-scale image recognition. arXiv 2014,
arXiv:1409.1556.
10. Szegedy, C.; Liu, W.; Jia, Y.; Sermanet, P.; Reed, S.; Anguelov, D.; Erhan, D.; Vanhoucke, V.; Rabinovich, A.
Going deeper with convolutions. In Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition, Boston, MA, USA, 7–12 June 2015; pp. 1–9.
11. He, K.; Zhang, X.; Ren, S.; Sun, J. Deep residual learning for image recognition. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition, Las Vegas, NV, USA, 27–30 June 2016; pp. 770–778.
12. Canziani, A.; Paszke, A.; Culurciello, E. An analysis of deep neural network models for practical applications.
arXiv 2016, arXiv:1605.07678.
Electronics 2019, 8, 292 53 of 66
13. Zweig, G. Classification and recognition with direct segment models. In Proceedings of the 2012 IEEE
International Conference on Acoustics, Speech and Signal Processing (ICASSP), Kyoto, Japan, 25–30 March
2012; pp. 4161–4164.
14. He, Y.; Fosler-Lussier, E. Efficient segmental conditional random fields for one-pass phone recognition. In
Proceedings of the Thirteenth Annual Conference of the International Speech Communication Association,
Portland, OR, USA, 9–13 September 2012.
15. Abdel-Hamid, O.; Deng, L.; Yu, D.; Jiang, H. Deep segmental neural networks for speech recognition.
Interspeech 2013, 36, 70.
16. Tang, H.; Wang, W.; Gimpel, K.; Livescu, K. Discriminative segmental cascades for feature-rich phone
recognition. In Proceedings of the 2015 IEEE Workshop on Automatic Speech Recognition and Understanding
(ASRU), Scottsdale, AZ, USA, 13–17 December 2015; pp. 561–568.
17. Song, W.; Cai, J. End-to-End Deep Neural Network for Automatic Speech Recognition. 1. (Errors: 21.1), 2015.
Available online: https://cs224d.stanford.edu/reports/SongWilliam.pdf (accessed on 17 January 2018).
18. Deng, L.; Abdel-Hamid, O.; Yu, D. A deep convolutional neural network using heterogeneous pooling
for trading acoustic invariance with phonetic confusion. In Proceedings of the 2013 IEEE International
Conference on Acoustics, Speech and Signal Processing (ICASSP), Vancouver, BC, Canada, 26–31 May 2013;
pp. 6669–6673.
19. Graves, A.; Mohamed, A.-R.; Hinton, G. Speech recognition with deep recurrent neural networks. In
Proceedings of the 2013 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP),
Vancouver, BC, Canada, 26–31 May 2013; pp. 6645–6649.
20. Zhang, Y.; Pezeshki, M.; Brakel, P.; Zhang, S.; Bengio, C.L.Y.; Courville, A. Towards end-to-end speech
recognition with deep convolutional neural networks. arXiv 2017, arXiv:1701.02720.
21. Deng, L.; Platt, J. Ensemble deep learning for speech recognition. In Proceedings of the Fifteenth Annual
Conference of the International Speech Communication Association, Singapore, 14–18 September 2014.
22. Chorowski, J.K.; Bahdanau, D.; Serdyuk, D.; Cho, K.; Bengio, Y. Attention-based models for speech
recognition. In Advances in Neural Information Processing Systems; MIT Press: Cambridge, MA, USA, 2015;
pp. 577–585.
23. Lu, L.; Kong, L.; Dyer, C.; Smith, N.A.; Renals, S. Segmental recurrent neural networks for end-to-end speech
recognition. arXiv 2016, arXiv:1603.00223.
24. Van Essen, B.; Kim, H.; Pearce, R.; Boakye, K.; Chen, B. LBANN: Livermore big artificial neural network
HPC toolkit. In Proceedings of the Workshop on Machine Learning in High-Performance Computing
Environments, Austin, TX, USA, 15–20 November 2015; p. 5.
25. Li, Y.; Yu, R.; Shahabi, C.; Liu, Y. Graph Convolutional Recurrent Neural Network: Data-Driven Traffic
Forecasting. arXiv 2017, arXiv:1707.01926.
26. Md, Z.A.; Aspiras, T.; Taha, T.M.; Asari, V.K.; Bowen, T.J. Advanced deep convolutional neural network
approaches for digital pathology image analysis: A comprehensive evaluation with different use cases. In
Proceedings of the Pathology Visions 2018, San Diego, CA, USA, 4–6 November 2018.
27. Long, J.; Shelhamer, E.; Darrell, T. Fully convolutional networks for semantic segmentation. In Proceedings
of the IEEE Conference on Computer Vision and Pattern Recognition, Boston, MA, USA, 7–12 June 2015;
pp. 3431–3440.
28. Alom, M.Z.; Yakopcic, C.; Taha, T.M.; Asari, V.K. Nuclei Segmentation with Recurrent Residual Convolutional
Neural Networks based U-Net (R2U-Net). In Proceedings of the NAECON 2018-IEEE National Aerospace
and Electronics Conference, Dayton, OH, USA, 23–26 July 2018; pp. 228–233.
29. Alom, M.Z.; Yakopcic, C.; Taha, T.M.; Asari, V.K. Microscopic Blood Cell Classification Using Inception
Recurrent Residual Convolutional Neural Networks. In Proceedings of the NAECON 2018-IEEE National
Aerospace and Electronics Conference, Dayton, OH, USA, 23–26 July 2018; pp. 222–227.
30. Chen, X.-W.; Lin, X. Big Data Deep Learning: Challenges and Perspectives. IEEE Access 2014, 2, 514–525.
[CrossRef]
31. Zhou, Z.-H.; Chawla, N.V.; Jin, Y.; Williams, G.J. Big data opportunities and challenges: Discussions from
data analytics perspectives. IEEE Comput. Intell. Mag. 2014, 9, 62–74. [CrossRef]
32. Najafabadi, M.M.; Villanustre, F.; Khoshgoftaar, T.M.; Seliya, N.; Wald, R.; Muharemagic, E. Deep learning
applications and challenges in big data analytics. J. Big Data 2015, 2, 1. [CrossRef]
Electronics 2019, 8, 292 54 of 66
33. Goodfellow, I.; Pouget-Abadie, J.; Mirza, M.; Xu, B.; Warde-Farley, D.; Ozair, S.; Courville, A.; Bengio, Y.
Generative adversarial nets. In Advances in Neural Information Processing Systems; The MIT Press: Cambridge,
MA, USA, 2014; pp. 2672–2680.
34. Kaiser, L.; Gomez, A.N.; Shazeer, N.; Vaswani, A.; Parmar, N.; Jones, L.; Uszkoreit, J. One model to learn
them all. arXiv 2017, arXiv:1706.05137.
35. Collobert, R.; Weston, J. A unified architecture for natural language processing: Deep neural networks with
multitask learning. In Proceedings of the 25th International Conference on Machine Learning, Helsinki,
Finland, 5–9 July 2008; pp. 160–167.
36. Johnson, M.; Schuster, M.; Le, Q.V.; Krikun, M.; Wu, Y.; Chen, Z.; Thorat, N.; Viégas, F.; Wattenberg, M.;
Corrado, G.; et al. Google’s multilingual neural machine translation system: Enabling zero-shot translation.
Trans. Assoc. Comput. Linguist. 2017, 5, 339–351. [CrossRef]
37. Argyriou, A.; Evgeniou, T.; Pontil, M. Multi-task feature learning. In Advances in Neural Information Processing
Systems; The MIT Press: Cambridge, MA, USA, 2007; pp. 41–48.
38. Singh, K.; Gupta, G.; Vig, L.; Shroff, G.; Agarwal, P. Deep Convolutional Neural Networks for Pairwise
Causality. arXiv 2017, arXiv:1701.00597.
39. Yu, H.; Wang, J.; Huang, Z.; Yang, Y.; Xu, W. Video paragraph captioning using hierarchical recurrent neural
networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas,
NV, USA, 27–30 June 2016; pp. 4584–4593.
40. Kim, T.; Cha, M.; Kim, H.; Lee, J.K.; Kim, J. Learning to discover cross-domain relations with generative
adversarial networks. arXiv 2017, arXiv:1703.05192.
41. Reed, S.; Akata, Z.; Yan, X.; Logeswaran, L.; Schiele, B.; Lee, H. Generative adversarial text to image synthesis.
arXiv 2016, arXiv:1605.05396.
42. Deng, L.; Yu, D. Deep learning: Methods and applications. Found. Trends Signal Process. 2014, 7, 197–387.
[CrossRef]
43. Gu, J.; Wang, Z.; Kuen, J.; Ma, L.; Shahroudy, A.; Shuai, B.; Liu, T.; Wang, X.; Wang, G.; Cai, J.; et al. Recent
advances in convolutional neural networks. arXiv 2015, arXiv:1512.07108.
44. Sze, V.; Chen, Y.; Yang, T.; Emer, J.S. Efficient processing of deep neural networks: A tutorial and survey.
Proc. IEEE 2017, 105, 2295–2329. [CrossRef]
45. Kwon, D.; Kim, H.; Kim, J.; Suh, S.C.; Kim, I.; Kim, K.J. A survey of deep learning-based network anomaly
detection. Cluster Comput. 2017, 1–13. [CrossRef]
46. Li, Y. Deep reinforcement learning: An overview. arXiv 2017, arXiv:1701.07274.
47. Kober, J.; Bagnell, J.A.; Peters, J. Reinforcement learning in robotics: A survey. Int. J. Robot. Res. 2013, 32,
1238–1274. [CrossRef]
48. Pan, S.J.; Yang, Q. A survey on transfer learning. IEEE Trans. Knowl. Data Eng. 2010, 22, 1345–1359. [CrossRef]
49. Schuman, C.D.; Potok, T.E.; Patton, R.M.; Birdwell, J.D.; Dean, M.E.; Rose, G.S.; Plank, J.S. A survey of
neuromorphic computing and neural networks in hardware. arXiv 2017, arXiv:1705.06963.
50. McCulloch, W.S.; Pitts, W. A logical calculus of the ideas immanent in nervous activity. Bull. Math. Biophys.
1943, 5, 115–133. [CrossRef]
51. Rosenblatt, F. The perceptron: A probabilistic model for information storage and organization in the brain.
Psychol. Rev. 1958, 65, 386. [CrossRef] [PubMed]
52. Minsky, M.; Papert, S.A. Perceptrons: An Introduction to Computational Geometry; MIT Press: Cambridge, MA,
USA, 2017.
53. Ackley, D.H.; Hinton, G.E.; Sejnowski, T.J. A learning algorithm for Boltzmann machines. Cogn. Sci. 1985, 9,
147–169. [CrossRef]
54. Fukushima, K. Neocognitron: A hierarchical neural network capable of visual pattern recognition.
Neural Netw. 1988, 1, 119–130. [CrossRef]
55. LeCun, Y.; Bottou, L.; Bengio, Y.; Haffner, P. Gradient-based learning applied to document recognition.
Proc. IEEE 1998, 86, 2278–2324. [CrossRef]
56. Hinton, G.E.; Osindero, S.; Teh, Y.-W. A fast learning algorithm for deep belief nets. Neural Comput. 2006, 18,
1527–1554. [CrossRef] [PubMed]
57. Hinton, G.E.; Salakhutdinov, R.R. Reducing the dimensionality of data with neural networks. Science 2006,
Electronics 2019, 8, 292 55 of 66
58. Bottou, L. Stochastic gradient descent tricks. In Neural Networks: Tricks of the Trade; Springer: Berlin/Heidelberg,
Germany, 2012; pp. 421–436.
59. Rumelhart, D.E.; Hinton, G.E.; Williams, R.J. Learning representations by back-propagating errors.
Cogn. Model. 1988, 5, 1. [CrossRef]
60. Sutskever, I.; Martens, J.; Dahl, G.; Hinton, G. On the importance of initialization and momentum in deep
learning. Int. Conf. Mach. Learning. 2013, 28, 1139–1147.
61. Yoshua, B.; Lamblin, P.; Popovici, D.; Larochelle, H. Greedy Layer-Wise Training of Deep Network. In
Advances in Neural Information Processing Systems 19 (NIPS 2006); MIT Press: Cambridge, MA, USA, 2007;
pp. 153–160.
62. Erhan, D.; Manzagol, P.; Bengio, Y.; Bengio, S.; Vincent, P. The difficulty of training deep architectures and
the effect of unsupervised pre-training. Artif. Intell. Stat. 2009, 5, 153–160.
63. Mohamed, A.-R.; Dahl, G.E.; Hinton, G. Acoustic modeling using deep belief networks. IEEE Trans. Audio
Speech Lang. Process. 2012, 20, 14–22. [CrossRef]
64. Nair, V.; Hinton, G.E. Rectified linear units improve restricted boltzmann machines. In Proceedings of the
27th International Conference on Machine Learning (ICML-10), Haifa, Israel, 21–24 June 2010; pp. 807–814.
65. Vincent, P.; Larochelle, H.; Bengio, Y.; Manzagol, P. Extracting and composing robust features with denoising
autoencoders. In Proceedings of the Twenty-fifth International Conference on Machine Learning, Helsinki,
Finland, 5–9 July 2008; pp. 1096–1103.
66. Lin, M.; Chen, Q.; Yan, S. Network in network. arXiv 2013, arXiv:1312.4400.
67. Springenberg, J.T.; Dosovitskiy, A.; Brox, T.; Riedmiller, M. Striving for simplicity: The all convolutional net.
arXiv 2014, arXiv:1412.6806.
68. Huang, G.; Liu, Z.; van der Maaten, L.; Weinberger, K.Q. Densely connected convolutional networks. In
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, HI, USA, 21–26
July 2017; pp. 4700–4708.
69. Larsson, G.; Maire, M.; Shakhnarovich, G. FractalNet: Ultra-Deep Neural Networks without Residuals. arXiv
2016, arXiv:1605.07648.
70. Szegedy, C.; Ioffe, S.; Vanhoucke, V. Inception-v4, inception-resnet and the impact of residual connections on
learning. arXiv 2016, arXiv:1602.07261.
71. Szegedy, C.; Vanhoucke, V.; Ioffe, S.; Shlens, J.; Wojna, Z. Rethinking the inception architecture for computer
vision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, NV,
USA, 27–30 June 2016; pp. 2818–2826.
72. Zagoruyko, S.; Komodakis, N. Wide Residual Networks. arXiv 2016, arXiv:1605.07146.
73. Xie, S.; Girshick, R.; Dollár, P.; Tu, Z.; He, K. Aggregated residual transformations for deep neural networks.
arXiv 2016, arXiv:1611.05431.
74. Veit, A.; Wilber, M.J.; Belongie, S. Residual networks behave like ensembles of relatively shallow networks.
In Advances in Neural Information Processing Systems; MIT Press: Cambridge, MA, USA, 2016; pp. 550–558.
75. Abdi, M.; Nahavandi, S. Multi-Residual Networks: Improving the Speed and Accuracy of Residual Networks.
arXiv 2016, arXiv:1609.05672.
76. Zhang, X.; Li, Z.; Loy, C.C.; Lin, D. Polynet: A pursuit of structural diversity in very deep networks.
In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, HI, USA,
21–26 July 2017; pp. 718–726.
77. Alom, M.Z.; Hasan, M.; Yakopcic, C.; Taha, T.M.; Asari, V.K. Improved inception-residual convolutional
neural network for object recognition. arXiv 2017, arXiv:1712.09888.
78. Ioffe, S.; Szegedy, C. Batch normalization: Accelerating deep network training by reducing internal covariate
shift. arXiv 2015, arXiv:1502.03167.
79. Sabour, S.; Frosst, N.; Hinton, G.E. Dynamic routing between capsules. In Advances in Neural Information
Processing Systems (NIPS); MIT Press: Cambridge, MA, USA, 2017; pp. 3856–3866.
80. Ren, S.; He, K.; Girshick, R.; Sun, J. Faster r-cnn: Towards real-time object detection with region proposal
networks. In Advances in Neural Information Processing Systems; MIT Press: Cambridge, MA, USA, 2015;
pp. 91–99.
81. Chollet, F. Xception: Deep Learning with Depthwise Separable Convolutions. arXiv 2016, arXiv:1610.02357.
82. Liang, M.; Hu, X. Recurrent convolutional neural network for object recognition. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition, Boston, MA, USA, 7–12 June 2015.
Electronics 2019, 8, 292 56 of 66
83. Alom, M.Z.; Hasan, M.; Yakopcic, C.; Taha, T.M. Inception Recurrent Convolutional Neural Network for
Object Recognition. arXiv 2017, arXiv:1704.07709.
84. Li, Y.; Ouyang, W.; Wang, X.; Tang, X. Vip-cnn: Visual phrase guided convolutional neural network. In
Proceedings of the 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Honolulu,
HI, USA, 21–26 July 2017; pp. 7244–7253.
85. Bagherinezhad, H.; Rastegari, M.; Farhadi, A. LCNN: Lookup-based Convolutional Neural Network. arXiv
2016, arXiv:1611.06473.
86. Bansal, A.; Chen, X.; Russell, B.; Gupta, A.; Ramanan, D. Pixelnet: Representation of the pixels, by the pixels,
and for the pixels. arXiv 2017, arXiv:1702.06506.
87. Huang, G.; Sun, Y.; Liu, Z.; Sedra, D.; Weinberger, K.Q. Deep networks with stochastic depth. In European
Conference on Computer Vision; Springer: Cham, Switzerland, 2016; pp. 646–661.
88. Lee, C.-Y.; Xie, S.; Gallagher, P.; Zhang, Z.; Tu, Z. Deeply-supervised nets. In Proceedings of the Artificial
Intelligence and Statistics, San Diego, CA, USA, 9–12 May 2015; pp. 562–570.
89. Pezeshki, M.; Fan, L.; Brakel, P.; Courville, A.; Bengio, Y. Deconstructing the ladder network architecture. In
Proceedings of the International Conference on Machine Learning, New York, NY, USA, 20–22 June 2016;
pp. 2368–2376.
90. Rawat, W.; Wang, Z. Deep convolutional neural networks for image classification: A comprehensive review.
Neural Comput. 2017, 29, 2352–2449. [CrossRef] [PubMed]
91. Tzeng, E.; Hoffman, J.; Darrell, T.; Saenko, K. Simultaneous deep transfer across domains and tasks. In
Proceedings of the IEEE International Conference on Computer Vision, Las Condes, Chile, 11–18 December
2015; pp. 4068–4076.
92. Ba, J.; Caruana, R. Do deep nets really need to be deep? In Advances in Neural Information Processing Systems;
NIPS Proceedings; MIT Press: Cambridge, MA, USA, 2014.
93. Urban, G.; Geras, K.J.; Kahou, S.E.; Aslan, O.; Wang, S.; Caruana, R.; Mohamed, A.; Philipose, M.;
Richardson, M. Do deep convolutional nets really need to be deep and convolutional? arXiv 2016,
arXiv:1603.05691.
94. Romero, A.; Ballas, N.; Kahou, S.E.; Chassang, A.; Gatta, C.; Bengio, Y. Fitnets: Hints for thin deep nets. arXiv
2014, arXiv:1412.6550.
95. Mishkin, D.; Matas, J. All you need is a good init. arXiv 2015, arXiv:1511.06422.
96. Pandey, G.; Dukkipati, A. To go deep or wide in learning? arXiv 2014, arXiv:1402.5634.
97. Ratner, A.J.; de Sa, C.M.; Wu, S.; Selsam, D.; Ré, C. Data programming: Creating large training sets, quickly.
98. Aberger, C.R.; Lamb, A.; Tu, S.; Nötzli, A.; Olukotun, K.; Ré, C. Emptyheaded: A relational engine for graph
processing. ACM Trans. Database Syst. 2017, 42, 20. [CrossRef]
99. Iandola, F.N.; Han, S.; Moskewicz, M.W.; Ashraf, K.; Dally, W.J.; Keutzer, K. Squeezenet: Alexnet-level
accuracy with 50x fewer parameters and <0.5 mb model size. arXiv 2016, arXiv:1602.07360.
100. Han, S.; Mao, H.; Dally, W.J. Deep compression: Compressing deep neural network with pruning, trained
quantization and huffman coding. arXiv 2015, arXiv:1510.00149.
101. Niepert, M.; Ahmed, M.; Kutzkov, K. Learning Convolutional Neural Networks for Graphs. arXiv 2016,
arXiv:1605.05273.
102. Awesome Deep Vision. Available online: https://github.com/kjw0612/awesome-deep-vision (accessed on
17 January 2018).
103. Jia, X.; Xu, X.; Cai, B.; Guo, K. Single Image Super-Resolution Using Multi-Scale Convolutional Neural
Network. In Pacific Rim Conference on Multimedia; Springer: Cham, Switzerland, 2017; pp. 149–157.
104. Ahn, B.; Cho, N.I. Block-Matching Convolutional Neural Network for Image Denoising. arXiv 2017,
arXiv:1704.00524.
105. Ma, S.; Liu, J.; Chen, C.W. A-Lamp: Adaptive Layout-Aware Multi-Patch Deep Convolutional Neural
Network for Photo Aesthetic Assessment. arXiv 2017, arXiv:1704.00248.
106. Cao, X.; Zhou, F.; Xu, L.; Meng, D.; Xu, Z.; Paisley, J. Hyperspectral Image Classification With Markov
Random Fields and a Convolutional Neural Network. IEEE Trans. Image Process. 2018, 27, 2354–2367.
[CrossRef] [PubMed]
Electronics 2019, 8, 292 57 of 66
107. De Vos, B.D.; Berendsen, F.F.; Viergever, M.A.; Staring, M.; Išgum, I. End-to-end unsupervised deformable
image registration with a convolutional neural network. In Deep Learning in Medical Image Analysis and
Multimodal Learning for Clinical Decision Support; Springer: Cham, Switzerland, 2017; pp. 204–212.
108. Wang, X.; Oxholm, G.; Zhang, D.; Wang, Y. Multimodal transfer: A hierarchical deep convolutional neural
network for fast artistic style transfer. In Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition, Honolulu, HI, USA, 21–26 July 2017; Volume 2, p. 7.
109. Babaee, M.; Dinh, D.T.; Rigoll, G. A deep convolutional neural network for background subtraction. arXiv
2017, arXiv:1702.01731.
110. Alom, M.Z.; Sidike, P.; Hasan, M.; Taha, T.M.; Asari, V.K. Handwritten Bangla Character Recognition Using
the State-of-the-Art Deep Convolutional Neural Networks. Comput. Intell. Neurosci. 2018, 2018, 6747098.
[CrossRef] [PubMed]
111. Alom, M.Z.; Awwal, A.A.S.; Lowe-Webb, R.; Taha, T.M. Optical beam classification using deep learning:
A comparison with rule-and feature-based classification. In Proceedings of the Optics and Photonics for
Information Processing XI, San Diego, CA, USA, 6–10 August 2017; Volume 10395.
112. Sidike, P.; Sagan, V.; Maimaitijiang, M.; Maimaitiyiming, M.; Shakoor, N.; Burken, J.; Mockler, T.; Fritschi, F.B.
dPEN: deep Progressively Expanded Network for mapping heterogeneous agricultural landscape using
WorldView-3 satellite imagery. Remote Sens. Environ. 2019, 221, 756–772. [CrossRef]
113. Alom, M.Z.; Alam, M.; Taha, T.M.; Iftekharuddin, K.M. Object recognition using cellular simultaneous
recurrent networks and convolutional neural network. In Proceedings of the 2017 International Joint
Conference on Neural Networks (IJCNN), Anchorage, AK, USA, 14–19 May 2017; pp. 2873–2880.
114. Ronao, C.A.; Cho, S.-B. Human activity recognition with smartphone sensors using deep learning neural
networks. Expert Syst. Appl. 2016, 59, 235–244. [CrossRef]
115. Yang, J.; Nguyen, M.N.; San, P.P.; Li, X.L.; Krishnaswamy, S. Deep convolutional neural networks on
multichannel time series for human activity recognition. In Proceedings of the Twenty-Fourth International
Joint Conference on Artificial Intelligence, Buenos Aires, Argentina, 25–31 July 2015.
116. Hammerla, N.Y.; Halloran, S.; Ploetz, T. Deep, convolutional, and recurrent models for human activity
recognition using wearables. arXiv 2016, arXiv:1604.08880.
117. Ordóñez, F.J.; Roggen, D. Deep convolutional and lstm recurrent neural networks for multimodal wearable
activity recognition. Sensors 2016, 16, 115. [CrossRef] [PubMed]
118. Rad, N.M.; Kia, S.M.; Zarbo, C.; van Laarhoven, T.; Jurman, G.; Venuti, P.; Marchiori, E.; Furlanello, C. Deep
learning for automatic stereotypical motor movement detection using wearable sensors in autism spectrum
disorders. Signal Process. 2018, 144, 180–191.
119. Ravi, D.; Wong, C.; Lo, B.; Yang, G. Deep learning for human activity recognition: A resource efficient
implementation on low-power devices. In Proceedings of the 2016 IEEE 13th International Conference
on Wearable and Implantable Body Sensor Networks (BSN), San Francisco, CA, USA, 14–17 June 2016;
pp. 71–76.
120. Alom, M.Z.; Yakopcic, C.; Taha, T.M.; Asari, V.K. Microscopic Nuclei Classification, Segmentation
and Detection with improved Deep Convolutional Neural Network (DCNN) Approaches. arXiv 2018,
arXiv:1811.03447.
121. Chen, L.-C.; Papandreou, G.; Kokkinos, I.; Murphy, K.; Yuille, A.L. Semantic image segmentation with deep
convolutional nets and fully connected crfs. arXiv 2014, arXiv:1412.7062.
122. Badrinarayanan, V.; Kendall, A.; Cipolla, R. Segnet: A deep convolutional encoder-decoder architecture for
image segmentation. arXiv 2015, arXiv:1511.00561.
123. Lin, G.; Milan, A.; Shen, C.; Reid, I. Refinenet: Multi-path refinement networks for high-resolution semantic
segmentation. In Proceedings of the 2017 IEEE Conference on Computer Vision and Pattern Recognition
(CVPR), Honolulu, HI, USA, 21–26 July 2017; pp. 5168–5177.
124. Zhao, H.; Shi, J.; Qi, X.; Wang, X.; Jia, J. Pyramid scene parsing network. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition (CVPR), Honolulu, HI, USA, 21–26 July 2017;
pp. 2881–2890.
125. Chen, L.-C.; Papandreou, G.; Kokkinos, I.; Murphy, K.; Yuille, A.L. Deeplab: Semantic image segmentation
with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Trans. Pattern Anal.
Mach. Intell. 2018, 40, 834–848. [CrossRef] [PubMed]
Electronics 2019, 8, 292 58 of 66
126. Ronneberger, O.; Fischer, P.; Brox, T. U-net: Convolutional networks for biomedical image segmentation.
In International Conference on Medical Image Computing and Computer-Assisted Intervention; Springer: Cham,
Switzerland, 2015; pp. 234–241.
127. Alom, M.Z.; Hasan, M.; Yakopcic, C.; Taha, T.M.; Asari, V.K. Recurrent Residual Convolutional Neural
Network based on U-Net (R2U-Net) for Medical Image Segmentation. arXiv 2018, arXiv:1802.06955.
128. Girshick, R.; Donahue, J.; Darrell, T.; Malik, J. Rich feature hierarchies for accurate object detection and
semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition,
Columbus, OH, USA, 23–28 June 2014; pp. 580–587.
129. Lin, T.-Y.; Goyal, P.; Girshick, R.; He, K.; Dollár, P. Focal loss for dense object detection. In Proceedings of the
IEEE International Conference on Computer Vision, Venice, Italy, 22–29 October 2017; pp. 2980–2988.
130. Wang, X.; Shrivastava, A.; Gupta, A. A-fast-rcnn: Hard positive generation via adversary for object detection.
21–26 July 2017.
131. He, K.; Gkioxari, G.; Dollár, P.; Girshick, R. Mask r-cnn. In Proceedings of the 2017 IEEE International
Conference on Computer Vision (ICCV), Venice, Italy, 22–29 October 2017; pp. 2980–2988.
132. Redmon, J.; Divvala, S.; Girshick, R.; Farhadi, A. You only look once: Unified, real-time object detection.
In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, NV, USA,
27–30 June 2016; pp. 779–788.
133. Liu, W.; Anguelov, D.; Erhan, D.; Szegedy, C.; Reed, S.; Fu, C.-Y.; Berg, A.C. Ssd: Single shot multibox
detector. In European Conference on Computer Vision; Springer: Cham, Switzerland, 2016; pp. 21–37.
134. Hou, J.-C.; Wang, S.; Lai, Y.; Tsao, Y.; Chang, H.; Wang, H. Audio-Visual Speech Enhancement Using
Multimodal Deep Convolutional Neural Networks. arXiv 2017, arXiv:1703.10893.
135. Xu, Y.; Kong, Q.; Huang, Q.; Wang, W.; Plumbley, M.D. Convolutional gated recurrent neural network
incorporating spatial features for audio tagging. In Proceedings of the 2017 International Joint Conference
on Neural Networks (IJCNN), Anchorage, AK, USA, 14–19 May 2017; pp. 3461–3466.
136. Litjens, G.; Kooi, T.; Bejnordi, B.E.; Setio, A.A.A.; Ciompi, F.; Ghafoorian, M.; van der Laak, J.A.;
van Ginneken, B.; Sánchez, C.I. A survey on deep learning in medical image analysis. Med. Image Anal. 2017,
137. Zhang, Z.; Xie, Y.; Xing, F.; McGough, M.; Yang, L. Mdnet: A semantically and visually interpretable
medical image diagnosis network. In Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition, Honolulu, HI, USA, 21–26 July 2017; pp. 6428–6436.
138. Tran, P.V. A fully convolutional neural network for cardiac segmentation in short-axis MRI. arXiv 2016,
arXiv:1604.00494.
139. Tan, J.H.U.; Acharya, R.; Bhandary, S.V.; Chua, K.C.; Sivaprasad, S. Segmentation of optic disc, fovea and
retinal vasculature using a single convolutional neural network. J. Comput. Sci. 2017, 20, 70–79. [CrossRef]
140. Moeskops, P.; Viergever, M.A.; Mendrik, A.M.; de Vries, L.S.; Benders, M.J.N.L.; Išgum, I. Automatic
segmentation of MR brain images with a convolutional neural network. IEEE Trans. Med Imaging 2016, 35,
1252–1261. [CrossRef] [PubMed]
141. Alom, M.Z.; Yakopcic, C.; Taha, T.M.; Asari, V.K. Breast Cancer Classification from Histopathological Images
with Inception Recurrent Residual Convolutional Neural Network. arXiv 2018, arXiv:1811.04241.
142. LeCun, Y.; Bottou, L.; Orr, G. Efficient BackProp. In Neural Networks: Tricks of the Trade; Orr, G., Müller, K.,
Eds.; Lecture Notes in Computer Science; Springer: Berlin, Germany, 2012.
143. Glorot, X.; Bengio, Y. Understanding the difficulty of training deep feedforward neural networks. In
Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics, Sardinia, Italy,
13–15 May 2010; pp. 249–256.
144. He, K.; Zhang, X.; Ren, S.; Sun, J. Delving deep into rectifiers: Surpassing human-level performance on
imagenet classification. In Proceedings of the IEEE International Conference on Computer Vision, Las Condes,
Chile, 11–18 December 2015; pp. 1026–1034.
145. Vedaldi, A.; Lenc, K. Matconvnet: Convolutional neural networks for matlab. In Proceedings of the 23rd
ACM International Conference on Multimedia, Brisbane, Australia, 26–30 October 2015; pp. 689–692.
146. Laurent, C.; Pereyra, G.; Brakel, P.; Zhang, Y.; Bengio, Y. Batch normalized recurrent neural networks. In
Proceedings of the 2016 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP),
Shanghai, China, 20–25 March 2016; pp. 2657–2661.
Electronics 2019, 8, 292 59 of 66
147. Lavin, A.; Gray, S. Fast algorithms for convolutional neural networks. In Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition, Las Vegas, NV, USA, 27–30 June 2016; pp. 4013–4021.
148. Clevert, D.-A.; Unterthiner, T.; Hochreiter, S. Fast and accurate deep network learning by exponential linear
units (elus). arXiv 2015, arXiv:1511.07289.
149. Li, Y.; Fan, C.; Li, Y.; Wu, Q.; Ming, Y. Improving deep neural network with multiple parametric exponential
linear units. Neurocomputing 2018, 301, 11–24. [CrossRef]
150. Jin, X.; Xu, C.; Feng, J.; Wei, Y.; Xiong, J.; Yan, S. Deep Learning with S-Shaped Rectified Linear Activation
Units. AAAI 2016, 3, 2–3.
151. Xu, B.; Wang, N.; Chen, T.; Li, M. Empirical evaluation of rectified activations in convolutional network.
arXiv 2015, arXiv:1505.00853.
152. He, K.; Zhang, X.; Ren, S.; Sun, J. Spatial pyramid pooling in deep convolutional networks for visual
recognition. In European Conference on Computer Vision; Springer: Cham, Switzerland, 2014; pp. 346–361.
153. Yoo, D.; Park, S.; Lee, J.; Kweon, I.S. Multi-scale pyramid pooling for deep convolutional representation. In
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, Boston, MA,
USA, 7–12 June 2015; pp. 71–80.
154. Graham, B. Fractional max-pooling. arXiv 2014, arXiv:1412.6071.
155. Lee, C.-Y.; Gallagher, P.W.; Tu, Z. Generalizing pooling functions in convolutional neural networks: Mixed,
gated, and tree. In Proceedings of the Artificial Intelligence and Statistics, Cadiz, Spain, 9–11 May 2016;
pp. 464–472.
156. Hinton, G.E.; Srivastava, N.; Krizhevsky, A.; Sutskever, I.; Salakhutdinov, R.R. Improving neural networks
by preventing co-adaptation of feature detectors. arXiv 2012, arXiv:1207.0580.
157. Srivastava, N.; Hinton, G.; Krizhevsky, A.; Sutskever, I.; Salakhutdinov, R. Dropout: A simple way to prevent
neural networks from overfitting. J. Mach. Learn. Res. 2014, 15, 1929–1958.
158. Wan, L.; Zeiler, M.; Zhang, S.; le Cun, Y.; Fergus, R. Regularization of neural networks using dropconnect.
In Proceedings of the International Conference on Machine Learning, Atlanta, GA, USA, 16–21 June 2013;
pp. 1058–1066.
159. Bulò, S.R.; Porzi, L.; Kontschieder, P. Dropout distillation. In Proceedings of the International Conference on
Machine Learning, New York, NY, USA, 20–22 June 2016; pp. 99–107.
160. Ruder, S. An overview of gradient descent optimization algorithms. arXiv 2016, arXiv:1609.04747.
161. Le, Q.V.; Ngiam, J.; Coates, A.; Lahiri, A.; Prochnow, B.; Ng, A.Y. On optimization methods for deep learning.
In Proceedings of the 28th International Conference on International Conference on Machine Learning,
Bellevue, WA, USA, 28 June –2 July 2011; pp. 265–272.
162. Koushik, J.; Hayashi, H. Improving stochastic gradient descent with feedback. arXiv 2016, arXiv:1611.01505.
163. Sathasivam, S.; Abdullah, W.A. Logic learning in Hopfield networks. arXiv 2008, arXiv:0804.4075.
164. Elman, J.L. Finding structure in time. Cogn. Sci. 1990, 14, 179–211. [CrossRef]
165. Jordan, M.I. Serial order: A parallel distributed processing approach. Adv. Psychol. 1997, 121, 471–495.
166. Hochreiter, S.; Bengio, Y.; Frasconi, P.; Schmidhuber, J. Gradient Flow in Recurrent Nets: The Difficulty of
Learning Long-Term Dependencies; IEEE Press: New York, NY, USA, 2001.
167. Schmidhuber, J. Habilitation Thesis: Netzwerkarchitekturen, Zielfunktionen und Kettenregel (Network
architectures, objective functions, and chain rule). Ph.D. Thesis, Technische Universität München, München,
Germany, 15 April 1993.
168. Gers, F.A.; Schmidhuber, J. Recurrent nets that time and count. In Proceedings of the IEEE-INNS-ENNS
International Joint Conference on Neural Networks, Como, Italy, 24–27 July 2000; Volume 3.
169. Gers, F.A.; Schraudolph, N.N.; Schmidhuber, J. Learning precise timing with LSTM recurrent networks.
J. Mach. Learn. Res. 2002, 3, 115–143.
170. Socher, R.; Lin, C.C.; Manning, C.; Ng, A.Y. Parsing natural scenes and natural language with recursive
neural networks. In Proceedings of the 28th International Conference on Machine Learning (ICML-11),
Bellevue, WA, USA, 28 June–2 July 2011; pp. 129–136.
171. Mikolov, T.; Karafiát, M.; Burget, L.; Černocký, J.; Khudanpur, S. Recurrent neural network based language
model. In Proceedings of the Eleventh Annual Conference of the International Speech Communication
Association. Makuhari, Chiba, Japan, 26–30 September 2010; Volume 2.
Electronics 2019, 8, 292 60 of 66
172. Xingjian, S.H.I.; Chen, Z.; Wang, H.; Yeung, D.; Wong, W.; Woo, W. Convolutional LSTM network: A machine
learning approach for precipitation nowcasting. In Advances in Neural Information Processing Systems (NIPS);
NIPS Proceedings; MIT Press: Cambridge, MA, USA, 2015; pp. 802–810.
173. Chung, J.; Gulcehre, C.; Cho, K.; Bengio, Y. Empirical evaluation of gated recurrent neural networks on
sequence modeling. arXiv 2014, arXiv:1412.3555.
174. Jozefowicz, R.; Zaremba, W.; Sutskever, I. An empirical exploration of recurrent network architectures. In
Proceedings of the 32nd International Conference on Machine Learning (ICML-15), Lille, France, 6–11 July 2015.
175. Yao, K.; Cohn, T.; Vylomova, K.; Duh, K.; Dyer, C. Depth-gated recurrent neural networks. arXiv 2015,
arXiv:1508.03790.
176. Koutnik, J.; Greff, K.; Gomez, F.; Schmidhuber, J. A clockwork rnn. arXiv 2014, arXiv:1402.3511.
177. Greff, K.; Srivastava, R.K.; Koutník, J.; Steunebrink, B.R.; Schmidhuber, J. LSTM: A search space odyssey.
IEEE Trans. Neural Netw. Learn. Syst. 2017, 28, 2222–2232. [CrossRef] [PubMed]
178. Karpathy, A.; Li, F.-F. Deep visual-semantic alignments for generating image descriptions. In Proceedings of
the IEEE Conference on Computer Vision and Pattern Recognition, Boston, MA, USA, 7–12 June 2015.
179. Mikolov, T.; Chen, K.; Corrado, G.; Dean, J. Efficient estimation of word representations in vector space.
arXiv 2013, arXiv:1301.3781.
180. Goldberg, Y.; Levy, O. word2vec Explained: Deriving Mikolov et al.’s negative-sampling word-embedding
method. arXiv 2014, arXiv:1402.3722.
181. Kunihiko, F. Neural network model for selective attention in visual pattern recognition and associative recall.
Appl. Opt. 1987, 26, 4985–4992.
182. Xu, K.; Ba, J.; Kiros, R.; Cho, K.; Courville, A.; Salakhudinov, R.; Zemel, R.; Bengio, Y. Show, attend and tell:
Neural image caption generation with visual attention. In Proceedings of the International Conference on
Machine Learning, Lille, France, 6–11 July 2015; pp. 2048–2057.
183. Qin, Y.; Song, D.; Chen, H.; Cheng, W.; Jiang, G.; Cottrell, G. A dual-stage attention-based recurrent neural
network for time series prediction. arXiv 2017, arXiv:1704.02971.
184. Xiong, C.; Merity, S.; Socher, R. Dynamic memory networks for visual and textual question answering. In
Proceedings of the International Conference on Machine Learning, New York, NY, USA, 20–22 June 2016.
185. Oord, A.v.d.; Kalchbrenner, N.; Kavukcuoglu, K. Pixel recurrent neural networks. arXiv 2016, arXiv:1601.06759.
186. Xue, W.; Nachum, I.B.; Pandey, S.; Warrington, J.; Leung, S.; Li, S. Direct estimation of regional wall
thicknesses via residual recurrent neural network. In International Conference on Information Processing in
Medical Imaging; Springer: Cham, Switzerland, 2017; pp. 505–516.
187. Tjandra, A.; Sakti, S.; Manurung, R.; Adriani, M.; Nakamura, S. Gated recurrent neural tensor network.
In Proceedings of the 2016 International Joint Conference on Neural Networks (IJCNN), Vancouver, BC,
Canada, 24–29 July 2016; pp. 448–455.
188. Wang, S.; Jing, J. Learning natural language inference with LSTM. arXiv 2015, arXiv:1512.08849.
189. Sutskever, I.; Vinyals, O.; Le, Q.V. Sequence to sequence learning with neural networks. In Advances in Neural
Information Processing Systems (NIPS); MIT Press: Cambridge, MA, USA, 2014; pp. 3104–3112.
190. Lakhani, V.A.; Mahadev, R. Multi-Language Identification Using Convolutional Recurrent Neural Network.
arXiv 2016, arXiv:1611.04010.
191. Längkvist, M.; Karlsson, L.; Loutfi, A. A review of unsupervised feature learning and deep learning for
time-series modeling. Pattern Recognit. Lett. 2014, 42, 11–24. [CrossRef]
192. Malhotra, P.; Vishnu, T.V.; Vig, L.; Agarwal, P.; Shroff, G. TimeNet: Pre-trained deep recurrent neural network
for time series classification. arXiv 2017, arXiv:1706.08838.
193. Soltau, H.; Liao, H.; Sak, H. Neural speech recognizer: Acoustic-to-word LSTM model for large vocabulary
speech recognition. arXiv 2016, arXiv:1610.09975.
194. Sak, H.; Senior, A.; Beaufays, F. Long short-term memory recurrent neural network architectures for large
scale acoustic modeling. In Proceedings of the Fifteenth Annual Conference of the International Speech
Communication Association, Singapore, 14–18 September 2014.
195. Adavanne, S.; Pertilä, P.; Virtanen, T. Sound event detection using spatial features and convolutional recurrent
neural network. arXiv 2017, arXiv:1706.02291.
196. Chien, J.-T.; Misbullah, A. Deep long short-term memory networks for speech recognition. In Proceedings of
the 2016 10th International Symposium on Chinese Spoken Language Processing (ISCSLP), Tianjin, China,
17–20 October 2016.
Electronics 2019, 8, 292 61 of 66
197. Choi, E.; Schuetz, A.; Stewart, W.F.; Sun, J. Using recurrent neural network models for early detection of
heart failure onset. J. Am. Med Inform. Assoc. 2016, 24, 361–370. [CrossRef] [PubMed]
198. Azzouni, A.; Pujolle, G. A Long Short-Term Memory Recurrent Neural Network Framework for Network
Traffic Matrix Prediction. arXiv 2017, arXiv:1705.05690.
199. Olabiyi, O.; Martinson, E.; Chintalapudi, V.; Guo, R. Driver Action Prediction Using Deep (Bidirectional)
Recurrent Neural Network. arXiv 2017, arXiv:1706.02257.
200. Kim, B.D.; Kang, C.M.; Lee, S.H.; Chae, H.; Kim, J.; Chung, C.C.; Choi, J.W. Probabilistic vehicle trajectory
prediction over occupancy grid map via recurrent neural network. arXiv 2017, arXiv:1704.07049.
201. Richard, A.; Gall, J. A bag-of-words equivalent recurrent neural network for action recognition. Comput. Vis.
Image Underst. 2017, 156, 79–91. [CrossRef]
202. Bontemps, L.; McDermott, J.; Le-Khac, N.-H. Collective Anomaly Detection Based on Long Short-Term
Memory Recurrent Neural Networks. In International Conference on Future Data and Security Engineering;
Springer International Publishing: Cham, Switzerland, 2016.
203. Kingma, D.P.; Welling, M. Stochastic gradient VB and the variational auto-encoder. In Proceedings of the
Second International Conference on Learning Representations (ICLR), Banff, AB, Canada, 14–16 April 2014.
204. Ng, A. Sparse autoencoder. CS294A Lect. Notes 2011, 72, 1–19.
205. Vincent, P.; Larochelle, H.; Lajoie, I.; Bengio, Y.; Manzagol, P. Stacked denoising autoencoders: Learning
useful representations in a deep network with a local denoising criterion. J. Mach. Learn. Res. 2010, 11,
3371–3408.
206. Zhang, R.; Isola, P.; Efros, A.A. Split-brain autoencoders: Unsupervised learning by cross-channel prediction.
arXiv 2016, arXiv:1611.09842.
207. Lu, J.; Deshpande, A.; Forsyth, D. CDVAE: Co-embedding Deep Variational Auto Encoder for Conditional
Variational Generation. arXiv 2016, arXiv:1612.00132.
208. Chicco, D.; Sadowski, P.; Baldi, P. Deep Autoencoder Neural Networks for Gene Ontology Annotation
Predictions. In Proceedings of the 5th ACM Conference on Bioinformatics, Computational Biology, and
Health Informatics—BCB ’14, Niagara Falls, NY, USA, 2–4 August 2010; pp. 533–540.
209. Alom, M.Z.; Taha, T.M. Network Intrusion Detection for Cyber Security using Unsupervised Deep Learning
Approaches. In Proceedings of the Aerospace and Electronics Conference (NAECON), Dayton, OH, USA,
27–30 June 2017.
210. Song, C.; Liu, F.; Huang, Y.; Wang, L.; Tan, T. Auto-encoder based data clustering. In Iberoamerican Congress
on Pattern Recognition; Springer: Berlin/Heidelberg, Germany, 2013; pp. 117–124.
211. Ahmad, M.; Protasov, S.; Khan, A.M. Hyperspectral Band Selection Using Unsupervised Non-Linear Deep
Auto Encoder to Train External Classifiers. arXiv 2017, arXiv:1705.06920.
212. Freund, Y.; Haussler, D. Unsupervised learning of distributions on binary vectors using two layer networks.
213. Larochelle, H.; Bengio, Y. Classification using discriminative restricted Boltzmann machines. In Proceedings
of the 25th International Conference on Machine Learning, Helsinki, Finland, 5–9 July 2008.
214. Salakhutdinov, R.; Hinton, G.E. Deep Boltzmann machines. AISTATS 2009, 1, 3.
215. Alom, M.Z.; Bontupalli, V.R.; Taha, T.M. Intrusion detection using deep belief networks. In Proceedings of
the Aerospace and Electronics Conference (NAECON), Dayton, OH, USA, 16–19 June 2015.
216. Alom, M.Z.; Sidike, P.; Taha, T.M.; Asari, V.K. Handwritten bangla digit recognition using deep learning.
arXiv 2017, arXiv:1705.02680.
217. Albalooshi, F.A.; Sidike, P.; Sagan, V.; Albalooshi, Y.; Asari, V.K. Deep Belief Active Contours (DBAC) with Its
Application to Oil Spill Segmentation from Remotely Sensed Aerial Imagery. Photogramm. Eng. Remote Sens.
2018, 84, 451–458. [CrossRef]
218. Mao, X.; Li, Q.; Xie, H.; Lau, R.Y.K.; Wang, Z.; Smolley, S.P. Least squares generative adversarial networks. In
Proceedings of the IEEE International Conference on Computer Vision, Venice, Italy, 22–29 October 2017;
pp. 2794–2802.
219. Salimans, T.; Goodfellow, I.; Zaremba, W.; Cheung, V.; Radford, A.; Chen, X. Improved techniques for
training gans. arXiv 2016, arXiv:1606.03498.
220. Vondrick, C.; Pirsiavash, H.; Torralba, A. Generating videos with scene dynamics. In Advances in Neural
Information Processing Systems; MIT Press: Cambridge, MA, USA, 2016; pp. 613–621.
Electronics 2019, 8, 292 62 of 66
221. Radford, A.; Metz, L.; Chintala, S. Unsupervised representation learning with deep convolutional generative
222. Wang, X.; Gupta, A. Generative image modeling using style and structure adversarial networks. In European
Conference on Computer Vision; Springer: Cham, Switzerland, 2016.
223. Chen, X.; Duan, Y.; Houthooft, R.; Schulman, J.; Sutskever, I.; Abbeel, P. InfoGAN: Interpretable
representation learning by information maximizing generative adversarial nets. In Advances in Neural
Information Processing Systems; MIT Press: Cambridge, MA, USA, 2016.
224. Im, D.J.; Kim, C.D.; Jiang, H.; Memisevic, R. Generating images with recurrent adversarial networks. arXiv
2016, arXiv:1602.05110.
225. Isola, P.; Zhu, J.; Zhou, T.; Efros, A.A. Image-to-image translation with conditional adversarial networks.
arXiv 2017, arXiv:1611.07004.
226. Liu, M.-Y.; Tuzel, O. Coupled generative adversarial networks. In Advances in Neural Information Processing
Systems; MIT Press: Cambridge, MA, USA, 2016.
227. Donahue, J.; Krähenbühl, P.; Darrell, T. Adversarial feature learning. arXiv 2016, arXiv:1605.09782.
228. Berthelot, D.; Schumm, T.; Metz, L. Began: Boundary equilibrium generative adversarial networks. arXiv
2017, arXiv:1703.10717.
229. Martin, A.; Chintala, S.; Bottou, L. Wasserstein gan. arXiv 2017, arXiv:1701.07875.
230. Gulrajani, I.; Ahmed, F.; Arjovsky, M.; Dumoulin, V.; Courville, A.C. Improved training of wasserstein gans.
231. He, K.; Wang, Y.; Hopcroft, J. A powerful generative model using random weights for the deep image
representation. In Advances in Neural Information Processing Systems; MIT Press: Cambridge, MA, USA, 2016.
232. Kos, J.; Fischer, I.; Song, D. Adversarial examples for generative models. arXiv 2017, arXiv:1702.06832.
233. Zhao, J.; Mathieu, M.; LeCun, Y. Energy-based generative adversarial network. arXiv 2016, arXiv:1609.03126.
234. Park, N.; Anand, A.; Moniz, J.R.A.; Lee, K.; Chakraborty, T.; Choo, J.; Park, H.; Kim, Y. MMGAN: Manifold
Matching Generative Adversarial Network for Generating Images. arXiv 2017, arXiv:1707.08273.
235. Laloy, E.; Hérault, R.; Jacques, D.; Linde, N. Efficient training-image based geostatistical simulation and
inversion using a spatial generative adversarial neural network. arXiv 2017, arXiv:1708.04975.
236. Eghbal-zadeh, H.; Widmer, G. Probabilistic Generative Adversarial Networks. arXiv 2017, arXiv:1708.01886.
237. Fowkes, J.; Sutton, C. A Bayesian Network Model for Interesting Itemsets. In Joint European Conference on
Machine Learning and Knowledge Disco in Databases; Springer International Publishing: Cham, Switzerland, 2016.
238. Mescheder, L.; Nowozin, S.; Geiger, A. Adversarial variational bayes: Unifying variational autoencoders and
generative adversarial networks. arXiv 2017, arXiv:1701.04722.
239. Nowozin, S.; Cseke, B.; Tomioka, R. f-gan: Training generative neural samplers using variational divergence
minimization. In Advances in Neural Information Processing Systems; MIT Press: Cambridge, MA, USA, 2016.
240. Li, C.; Wand, M. Precomputed real-time texture synthesis with markovian generative adversarial networks.
In European Conference on Computer Vision; Springer International Publishing: Cham, Switzerland, 2016.
241. Du, C.; Zhu, J.; Zhang, B. Learning Deep Generative Models with Doubly Stochastic Gradient MCMC.
IEEE Trans. Neural Networks Learn. Syst. 2018, 29, 3084–3096. [CrossRef] [PubMed]
242. 242. Hoang, Quan, Tu Dinh Nguyen, Trung Le, and Dinh Phung. Multi-Generator Gernerative Adversarial
Nets. arXiv 2017, arXiv:1708.02556.
243. Bousmalis, K.; Silberman, N.; Dohan, D.; Erhan, D.; Krishnan, D. Unsupervised pixel-level domain adaptation
with generative adversarial networks. In Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition (CVPR), Honolulu, HI, USA, 21–26 July 2017; Volume 1, p. 7.
244. Kansky, K.; Silver, T.; Mély, D.A.; Eldawy, M.; Lázaro-Gredilla, M.; Lou, X.; Dorfman, N.; Sidor, S.; Phoenix, S.;
George, D. Schema networks: Zero-shot transfer with a generative causal model of intuitive physics. arXiv
2017, arXiv:1706.04317.
245. Ledig, C.; Theis, L.; Huszár, F.; Caballero, J.; Cunningham, A.; Acosta, A.; Aitken, A.; Tejani, A.; Totz, J.;
Wang, Z.; et al. Photo-realistic single image super-resolution using a generative adversarial network. arXiv
2016, arXiv:1609.04802.
246. Souly, N.; Spampinato, C.; Shah, M. Semi and Weakly Supervised Semantic Segmentation Using Generative
Adversarial Network. arXiv 2017, arXiv:1703.09695.
247. Dash, A.; Gamboa, J.C.B.; Ahmed, S.; Liwicki, M.; Afzal, M.Z. TAC-GAN-text conditioned auxiliary classifier
generative adversarial network. arXiv 2017, arXiv:1703.06412.
Electronics 2019, 8, 292 63 of 66
248. Zhang, H.; Dana, K. Multi-style Generative Network for Real-time Transfer. arXiv 2017, arXiv:1703.06953.
249. Zhang, H.; Sindagi, V.; Patel, V.M. Image De-raining Using a Conditional Generative Adversarial Network.
arXiv 2017, arXiv:1701.05957.
250. Serban, I.V.; Sordoni, A.; Bengio, Y.; Courville, A.C.; Pineau, J. Building End-To-End Dialogue Systems Using
Generative Hierarchical Neural Network Models. AAAI 2016, 16, 3776–3784.
251. Pascual, S.; Bonafonte, A.; Serrà, J. SEGAN: Speech Enhancement Generative Adversarial Network. arXiv
2017, arXiv:1703.09452.
252. Yang, L.-C.; Chou, S.-Z.; Yang, Y.-I. MidiNet: A convolutional generative adversarial network for
symbolic-domain music generation. In Proceedings of the 18th International Society for Music Information
Retrieval Conference (ISMIR’2017), Suzhou, China, 23–27 October 2017.
253. Yang, Q.; Yan, P.; Zhang, Y.; Yu, H.; Shi, Y.; Mou, X.; Kalra, M.K.; Zhang, Y.; Sun, L.; Wang, G. Low-dose
CT image denoising using a generative adversarial network with Wasserstein distance and perceptual loss.
IEEE Trans. Med. Imaging 2018, 37, 1348–1357. [CrossRef] [PubMed]
254. Rezaei, M.; Harmuth, K.; Gierke, W.; Kellermeier, T.; Fischer, M.; Yang, H.; Meinel, C. A conditional
adversarial network for semantic segmentation of brain tumor. In International MICCAI Brainlesion Workshop;
Springer: Cham, Switzerland, 2017; pp. 241–252.
255. Xue, Y.; Xu, T.; Zhang, H.; Long, L.R.; Huang, X. Segan: Adversarial network with multi-scale l 1 loss for
medical image segmentation. Neuroinformatics 2018, 16, 383–392. [CrossRef] [PubMed]
256. Mardani, M.; Gong, E.; Cheng, J.Y.; Vasanawala, S.; Zaharchuk, G.; Alley, M.; Thakur, N.; Han, S.; Dally, W.;
Pauly, J.M.; et al. Deep generative adversarial networks for compressed sensing automates MRI. arXiv 2017,
arXiv:1706.00051.
257. Choi, E.; Biswal, S.; Malin, B.; Duke, J.; Stewart, W.F.; Sun, J. Generating Multilabel Discrete Electronic Health
Records Using Generative Adversarial Networks. arXiv 2017, arXiv:1703.06490.
258. Esteban, C.; Hyland, S.L.; Rätsch, G. Real-valued (medical) time series generation with recurrent conditional
gans. arXiv 2017, arXiv:1706.02633.
259. Hayes, J.; Melis, L.; Danezis, G.; de Cristofaro, E. LOGAN: evaluating privacy leakage of generative models
using generative adversarial networks. arXiv 2017, arXiv:1705.07663.
260. Gordon, J.; Hernández-Lobato, J.M. Bayesian Semisupervised Learning with Deep Generative Models. arXiv
2017, arXiv:1706.09751.
261. Abbasnejad, M.E.; Shi, Q.; Abbasnejad, I.; van den Hengel, A.; Dick, A. Bayesian conditional generative
adverserial networks. arXiv 2017, arXiv:1706.05477.
262. Grnarova, P.; Levy, K.Y.; Lucchi, A.; Hofmann, T.; Krause, A. An online learning approach to generative
263. Li, Y.; Swersky, K.; Zemel, R. Generative moment matching networks. In Proceedings of the International
Conference on Machine Learning, Lille, France, 6–11 July 2015; pp. 1718–1727.
264. Li, C.-L.; Chang, W.; Cheng, Y.; Yang, Y.; Póczos, B. Mmd gan: Towards deeper understanding of moment
matching network. In Advances in Neural Information Processing Systems; MIT Press: Cambridge, MA, USA,
2017; pp. 2203–2213.
265. Nie, X.; Feng, J.; Xing, J.; Yan, S. Generative partition networks for multi-person pose estimation. arXiv 2017,
arXiv:1705.07422.
266. Saeedi, A.; Hoffman, M.D.; DiVerdi, S.J.; Ghandeharioun, A.; Johnson, M.J.; Adams, R.P. Multimodal
prediction and personalization of photo edits with deep generative models. arXiv 2017, arXiv:1704.04997.
267. Schlegl, T.; Seeböck, P.; Waldstein, S.M.; Schmidt-Erfurth, U.; Langs, G. Unsupervised anomaly detection
with generative adversarial networks to guide marker discovery. In International Conference on Information
Processing in Medical Imaging; Springer: Cham, Switzerland, 2017; pp. 146–157.
268. Liu, M.-Y.; Breuel, T.; Kautz, J. Unsupervised image-to-image translation networks. In Advances in Neural
269. Mehrotra, A.; Dukkipati, A. Generative Adversarial Residual Pairwise Networks for One Shot Learning.
arXiv 2017, arXiv:1703.08033.
270. Sordoni, A.; Galley, M.; Auli, M.; Brockett, C.; Ji, Y.; Mitchell, M.; Nie, J.; Gao, J.; Dolan, B. A neural network
approach to context-sensitive generation of conversational responses. arXiv 2015, arXiv:1506.06714.
271. Yin, J.; Jiang, X.; Lu, Z.; Shang, L.; Li, H.; Li, X. Neural generative question answering. arXiv 2015,
arXiv:1512.01337.
Electronics 2019, 8, 292 64 of 66
272. Oord, A.v.d.; Dieleman, S.; Zen, H.; Simonyan, K.; Vinyals, O.; Graves, A.; Kalchbrenner, N.; Senior, A.;
Kavukcuoglu, K. Wavenet: A generative model for raw audio. arXiv 2016, arXiv:1609.03499.
273. Chen, Y.; Li, J.; Xiao, H.; Jin, X.; Yan, S.; Feng, J. Dual path networks. In Advances in Neural Information
Processing Systems; MIT Press: Cambridge, MA, USA, 2017; pp. 4467–4475.
274. Mahmud, M.; Kaiser, M.S.; Hussain, A.; Vassanelli, S. Applications of deep learning and reinforcement
learning to biological data. IEEE Trans. Neural Netw. Learn. Syst. 2018, 29, 2063–2079. [CrossRef] [PubMed]
275. Goodfellow, I.; Bengio, Y.; Courville, A. Deep Learning; MIT Press: Cambridge, MA, USA, 2016.
276. Silver, D.; Huang, A.; Maddison, C.J.; Guez, A.; Sifre, L.; van den Driessche, G.; Schrittwieser, J. Mastering
the game of Go with deep neural networks and tree search. Nature 2016, 529, 484. [CrossRef] [PubMed]
277. Vinyals, O.; Ewalds, T.; Bartunov, S.; Georgiev, P.; Vezhnevets, A.S.; Yeo, M.; Makhzani, A.; Küttler, H.;
Agapiou, J.; Schrittwieser, J.; et al. Starcraft ii: A new challenge for reinforcement learning. arXiv 2017,
arXiv:1708.04782.
278. Koenig, S.; Simmons, R.G. Complexity Analysis of Real-Time Reinforcement Learning Applied to Finding Shortest
Paths in Deterministic Domains; Tech. Report, No. CMU-CS-93-106; Computer Science Department,
Carnegie-Mellon University: Pittsburgh PA, Decemver, 1992.
279. Silver, D.; Schrittwieser, J.; Simonyan, K.; Antonoglou, I.; Huang, A.; Guez, A.; Hubert, T.; Baker, L.; Lai, M.;
Bolton, A.; et al. Mastering the game of go without human knowledge. Nature 2017, 550, 354. [CrossRef]
[PubMed]
280. Schulman, J.; Levine, S.; Abbeel, P.; Jordan, M.I.; Moritz, P. Trust Region Policy Optimization. In Proceedings
of the 32nd International Conference on Machine Learning (ICML-15), Lille, France, 6–11 July 2015; Volume
37, pp. 1889–1897.
281. Levine, S.; Finn, C.; Darrell, T.; Abbeel, P. End-to-end training of deep visuomotor policies. J. Mach. Learn. Res.
2016, 17, 1334–1373.
282. Mnih, V.; Badia, A.P.; Mirza, M.; Graves, A.; Lillicrap, T.; Harley, T.; Silver, D.; Kavukcuoglu, K. Asynchronous
methods for deep reinforcement learning. In Proceedings of the International Conference on Machine
Learning, New York, NY, USA, 20–22 June 2016; pp. 1928–1937.
283. Arulkumaran, K.; Deisenroth, M.P.; Brundage, M.; Bharath, A.A. A brief survey of deep reinforcement
learning. arXiv 2017, arXiv:1708.05866.
284. Zhu, F.; Liao, P.; Zhu, X.; Yao, Y.; Huang, J. Cohesion-based online actor-critic reinforcement learning for
mhealth intervention. arXiv 2017, arXiv:1703.10039.
285. Zhu, F.; Guo, J.; Xu, Z.; Liao, P.; Yang, L.; Huang, J. Group-driven reinforcement learning for personalized
mhealth intervention. In International Conference on Medical Image Computing and Computer-Assisted
Intervention; Springer: Cham, Switzerland, 2018; pp. 590–598.
286. Steckelmacher, D.; Roijers, D.M.; Harutyunyan, A.; Vrancx, P.; Plisnier, H.; Nowé, A. Reinforcement
learning in POMDPs with memoryless options and option-observation initiation sets. In Proceedings
of the Thirty-Second AAAI Conference on Artificial Intelligence, New Orleans, LA, USA, 2–7 February 2018.
287. Hu, H.; Zhang, X.; Yan, X.; Wang, L.; Xu, Y. Solving a new 3d bin packing problem with deep reinforcement
learning method. arXiv 2017, arXiv:1708.05930.
288. Everitt, T.; Krakovna, V.; Orseau, L.; Hutter, M.; Legg, S. Reinforcement learning with a corrupted reward
channel. arXiv 2017, arXiv:1705.08417.
289. Wu, Y.; Mansimov, E.; Grosse, R.B.; Liao, S.; Ba, J. Scalable trust-region method for deep reinforcement
learning using kronecker-factored approximation. In Advances in Neural Information Processing Systems; MIT
Press: Cambridge, MA, USA, 2017; pp. 5279–5288.
290. Denil, M.; Agrawal, P.; Kulkarni, T.D.; Erez, T.; Battaglia, P.; de Freitas, N. Learning to perform physics
experiments via deep reinforcement learning. arXiv 2016, arXiv:1611.01843.
291. Hein, D.; Hentschel, A.; Runkler, T.; Udluft, S. Particle swarm optimization for generating interpretable
fuzzy reinforcement learning policies. Eng. Appl. Artif. Intell. 2017, 65, 87–98. [CrossRef]
292. Islam, R.; Henderson, P.; Gomrokchi, M.; Precup, D. Reproducibility of benchmarked deep reinforcement
learning tasks for continuous control. arXiv 2017, arXiv:1708.04133.
293. Inoue, T.; de Magistris, G.; Munawar, A.; Yokoya, T.; Tachibana, R. Deep reinforcement learning for high
precision assembly tasks. In Proceedings of the 2017 IEEE/RSJ International Conference on Intelligent
Robots and Systems (IROS), Vancouver, BC, Canada, 24–28 September 2017; pp. 819–825.
Electronics 2019, 8, 292 65 of 66
294. Li, K.; Burdick, J.W. Inverse Reinforcement Learning in Large State Spaces via Function Approximation.
arXiv 2017, arXiv:1707.09394.
295. Liu, N.; Li, Z.; Xu, J.; Xu, Z.; Lin, S.; Qiu, Q.; Tang, J.; Wang, Y. A hierarchical framework of cloud resource
allocation and power management using deep reinforcement learning. In Proceedings of the 2017 IEEE 37th
International Conference on Distributed Computing Systems (ICDCS), Atlanta, GA, USA, 5–8 June 2017;
pp. 372–382.
296. Cao, Q.; Lin, L.; Shi, Y.; Liang, X.; Li, G. Attention-aware face hallucination via deep reinforcement learning.
21–26 July 2017; pp. 690–698.
297. Kendall, A.; Gal, Y. What uncertainties do we need in bayesian deep learning for computer vision? In
Advances in Neural Information Processing Systems (NIPS); MIT Press: Cambridge, MA, USA, 2017.
298. Kendall, A.; Gal, Y.; Cipolla, R. Multi-task learning using uncertainty to weigh losses for scene geometry and
semantics. arXiv 2017, arXiv:1705.07115.
299. Google Photos labeled black people ‘gorillas’. Available online: https://www.usatoday.com/story/tech/
2015/07/01/google-apologizes-after-photos-identify-black-people-as-gorillas/29567465/ (accessed on
1 March 2019).
300. Gal, Y.; Ghahramani, Z. Bayesian convolutional neural networks with Bernoulli approximate variational
inference. arXiv 2015, arXiv:1506.02158.
301. Kumar, S.; Laumann, F.; Maurin, A.L.; Olsen, M.; Bayesian, M.L. Convolutional Neural Networks with
Variational Inference. arXiv 2018, arXiv:1704.02798.
302. Vladimirova, M.; Arbel, J.; Mesejo, P. Bayesian neural networks become heavier-tailed with depth. In
Proceedings of the Bayesian Deep Learning Workshop during the Thirty-Second Conference on Neural
Information Processing Systems (NIPS 2018), Montréal, QC, Canada, 7 December 2018.
303. Hu, S.X.; Champs-sur-Marne, F.; Moreno, P.G.; Lawrence, N.; Damianou, A. β-BNN: A Rate-Distortion
Perspective on Bayesian Neural Networks. In Proceedings of the Bayesian Deep Learning Workshop during
the Thirty-Second Conference on Neural Information Processing Systems (NIPS 2018), Montréal, QC, Canada,
7 December 2018.
304. Salvator, L.; Han, J.; Schroers, C.; Mandt, S. Video Compression through Deep Bayesian Learning Bayesian.
In Proceedings of the Deep Learning Workshop during the Thirty-Second Conference on Neural Information
Processing Systems (NIPS 2018), Montréal, QC, Canada, 7 December 2018.
305. Krishnan, R.; Subedar, M.; Tickoo, O. BAR: Bayesian Activity Recognition using variational inference. arXiv
2018, arXiv:1811.03305.
306. Chen, T.; Goodfellow, I.; Shlens, J. Net2net: Accelerating learning via knowledge transfer. arXiv 2015,
arXiv:1511.05641.
307. Ganin, Y.; Lempitsky, V. Unsupervised domain adaptation by backpropagation. arXiv 2014, arXiv:1409.7495.
308. Ganin, Y.; Ustinova, E.; Ajakan, H.; Germain, P.; Larochelle, H.; Laviolette, F.; Marchand, M.; Lempitsky, V.
Domain-adversarial training of neural networks. J. Mach. Learn. Res. 2016, 17, 2096–2130.
309. Taylor, M.E.; Stone, P. Transfer learning for reinforcement learning domains: A survey. J. Mach. Learn. Res.
2009, 10, 1633–1685.
310. McKeough, A. Teaching for Transfer: Fostering Generalization in Learning; Routledge: London, UK, 2013.
311. Raina, R.; Battle, A.; Lee, H.; Packer, B.; Ng, A.Y. Self-taught learning: transfer learning from unlabeled data.
In Proceedings of the 24th international conference on Machine learning, Corvallis, OR, USA, 20–24 June
2007; pp. 759–766.
312. Wenyuan, D.; Yang, Q.; Xue, G.; Yu, Y. Boosting for transfer learning. In Proceedings of the 24th International
Conference on Machine Learning, Corvallis, OR, USA, 20–24 June 2007; pp. 193–200.
313. Wu, Y.; Schuster, M.; Chen, Z.; Le, Q.V.; Norouzi, M.; Macherey, W.; Krikun, M.; Cao, Y.; Gao, Q.; Macherey, K.;
et al. Google’s neural machine translation system: Bridging the gap between human and machine translation.
arXiv 2016, arXiv:1609.08144.
314. Qiu, J.; Wang, J.; Yao, S.; Guo, K.; Li, B.; Zhou, E.; Yu, J.; Tang, T.; Xu, N.; Song, S.; et al. Going deeper
with embedded fpga platform for convolutional neural network. In Proceedings of the 2016 ACM/SIGDA
International Symposium on Field-Programmable Gate Arrays, Monterey, CA, USA, 21–23 February 2016;
pp. 26–35.
Electronics 2019, 8, 292 66 of 66
315. He, K.; Sun, J. Convolutional neural networks at constrained time cost. In Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition, Boston, MA, USA, 7–12 June 2015; pp. 5353–5360.
316. Lin, Z.; Courbariaux, M.; Memisevic, R.; Bengio, Y. Neural networks with few multiplications. arXiv 2015,
arXiv:1510.03009.
317. Courbariaux, M.; David, J.-E.; Bengio, Y. Training deep neural networks with low precision multiplications.
arXiv 2014, arXiv:1412.7024.
318. Courbariaux, M.; Bengio, Y.; David, J.-P. Binaryconnect: Training deep neural networks with binary weights
during propagations. In Advances in Neural Information Processing Systems; MIT Press: Cambridge, MA, USA, 2015.
319. Hubara, I.; Soudry, D.; El Yaniv, R. Binarized Neural Networks. arXiv 2016, arXiv:1602.02505.
320. Kim, M.; Smaragdis, P. Bitwise neural networks. arXiv 2016, arXiv:1601.06071.
321. Dettmers, T. 8-Bit Approximations for Parallelism in Deep Learning. arXiv 2015, arXiv:1511.04561.
322. Gupta, S.; Agrawal, A.; Gopalakrishnan, K.; Narayanan, P. Deep learning with limited numerical precision.
In Proceedings of the International Conference on Machine Learning, Lille, France, 6–11 July 2015;
pp. 1737–1746.
323. Zhou, S.; Wu, Y.; Ni, Z.; Zhou, X.; Wen, H.; Zou, Y. Dorefa-net: Training low bitwidth convolutional neural
networks with low bitwidth gradients. arXiv 2016, arXiv:1606.06160.
324. Merolla, P.A.; Arthur, J.V.; Alvarez-Icaza, R.; Cassidy, A.S.; Sawada, J.; Akopyan, F.; Jackson, B.L.; Imam, N.;
Guo, C.; Nakamura, Y.; et al. A million spiking-neuron integrated circuit with a scalable communication
network and interface. Science 2014, 345, 668–673. [CrossRef] [PubMed]
325. Steven, K.E.; Merolla, P.A.; Arthur, J.V.; Cassidy, A.S. Convolutional networks for fast, energy-efficient
neuromorphic computing. Proc. Natl. Acad. Sci. USA 2016, 27, 201604850.
326. Zidan, M.A.; Strachan, J.P.; Lu, W.D. The future of electronics based on memristive systems. Nat. Electron.
2018, 1, 22. [CrossRef]
327. Chen, Y.-H.; Krishna, T.; Emer, J.S.; Sze, V. Eyeriss: An energy-efficient reconfigurable accelerator for deep
convolutional neural networks. IEEE J. Solid-State Circuits 2017, 52, 127–138. [CrossRef]
328. Chen, Y.; Luo, T.; Liu, S.; Zhang, S.; He, L.; Wang, J.; Li, L.; Chen, T.; Xu, Z.; Sun, N.; et al. Dadiannao: A
machine-learning supercomputer. In Proceedings of the 47th Annual IEEE/ACM International Symposium
on Microarchitecture, Cambridge, UK, 13–17 December 2014; pp. 609–622.
329. Jouppi, N.P.; Young, C.; Patil, N.; Patterson, D.; Agrawal, G.; Bajwa, R.; Bates, S.; Bhatia, S.; Boden, N.;
Borchers, A.; et al. In-datacenter performance analysis of a tensor processing unit. In Proceedings of the 2017
ACM/IEEE 44th Annual International Symposium on Computer Architecture (ISCA), Toronto, ON, Canada,
24–28 June 2017; pp. 1–12.
330. Han, S.; Liu, X.; Mao, H.; Pu, J.; Pedram, A.; Horowitz, M.A.; Dally, W.J. EIE: Efficient inference engine
on compressed deep neural network. In Proceedings of the 2016 ACM/IEEE 43rd Annual International
Symposium on Computer Architecture (ISCA), Seoul, Korea, 18–22 June 2016; pp. 243–254.
331. Zhang, X.; Zou, J.; Ming, X.; He, K.; Sun, J. Efficient and accurate approximations of nonlinear convolutional
networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Boston, MA,
USA, 7–12 June 2015; pp. 1984–1992.
332. Novikov, A.; Podoprikhin, D.; Osokin, A.; Vetrov, D.P. Tensorizing neural networks. In Advances in Neural
333. Zhu, C.; Han, S.; Mao, H.; Dally, W.J. Trained ternary quantization. arXiv 2016, arXiv:1612.01064.
© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).

Electronics 08 00292 v3 PDF

Uploaded by

Copyright:

Available Formats

Electronics 08 00292 v3 PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Electronics 08 00292 v3 PDF

Uploaded by

Copyright:

Available Formats

electronics

Electronics 2019, 8, 292; doi:10.3390/electronics8030292 www.mdpi.com/journal/electronics

1.1.4. Deep Reinforcement Learning (RL)

1.2. Feature Learning

1.1.1. Deep Supervised Learning

1.1.2. Deep Semi-supervised Learning

1.1.3. Deep Unsupervised Learning

1.1.4. Deep Reinforcement Learning (RL)

1.2. Feature Learning

Table 1. Different feature learning approaches.

Approaches Learning Steps

1.3. Why and When to apply DL

• Absence of a human expert (navigation on Mars)

The experimental results on ImageNet-2012

AlexNet[7] Clarifia[8] VGG-16 [9] GoogLeNet- ResNet- Human

1.4. The State-of-the-art Performance of RNN[23]

CDNN w. Hater.s Pooling[18]

Phone error rate (PER) in percentage(%)

CDNN w. Hater.s Pooling[18]

DSC with 2nd pass[16]

Discriminative segmetal cascade (DSC)[16]

Deep Segmental NN[15]

First-pass SCRF [13]

Semantic segmentation [27]

Semantic segmentation [27]

Nuclei segmentation [28] Blood cell classification [29]

1.5. Why DL?

2. Deep Neural Network

2.1. The History of DNN

Figure 8. The history of deep learning development [50–57].

vt𝑣==γγv𝑣t−1 −−η𝜂∇ F (θt−1),),

2.6. Learning Rate (η )

2.7. Weight Decay

3. Convolutional Neural Network (CNN)

3.1.1. Convolutional Layer

x lj = f  ∑ xil −1 ∗ klij + blj , (8)

3.1.2. Sub-sampling Layer

3.1.3. Classification Layer

Feature maps after performing convolution and pooling operations.

3.1.4. Network Parameters and Required Memory for CNN

3.2. Popular CNN Architectures

Electronics 2019, 8, x FOR PEER REVIEW

3.2. Popular CNN Architectures

Figure 11. The architecture of LeNet.

3.2.4. Network in Network (NiN)

3.2.4. Network in Network (NiN)

Figure 14. Inception layer: Naive version.

3.2.6. GoogLeNet (2014)

Electronics 2019, 8, x FOR PEER REVIEW 17 of 67

Figure 15. Inception layer with dimension reduction.

3.2.7. Residual Network (ResNet in 2015)

Figure 18. A 4-layer Dense block with a growth rate of 𝑘 = 3.

Electronics 2019, 8, x FOR PEER REVIEW 20 of 67

3.4. Comparison of Different Models

3.5. Other DNN Models

3.6. Applications of CNNs

3.6.1. CNNs for Solving A Graph Problem

3.6.2. Image Processing and Computer Vision

3.6.3. Speech Processing