International Journal of Scientific Research in Science, Engineering and Technology (www.ijsrset.

com)
© 2020 IJSRSET | Volume 7 | Issue 2 | Print ISSN: 2395-1990 | Online ISSN : 2394-4099
DOI : https://doi.org/10.32628/IJSRSET207279
An Overview Studying of Deep Learning
Mya Sandar Kyin, Zaw Lin Oo, Khin Mar Cho
Faculty of Computer Science, University of Computer Studies, Taungoo, Pegu Regional Division, Myanmar

ABSTRACT

Deep learning is a subfield of machine learning however both drop under the broad category of artificial
intelligence. Deep learning is what powers the most human-like artificial intelligence that consents
computational models that are composed of multiple processing layers to learn representations of data with
multiple levels of abstraction. Deep learning is making major advances in solving problems hence categorized in
wider section of artificial intelligence. The main advantage of Deep Learning is to create an artificial neural
network that can learn and make intelligent decisions on its own and to process large numbers of features
makes deep learning very powerful when dealing with unstructured data.
Keywords: Deep Learning, Machine Learning

I. INTRODUCTION as they directly relate to, Machine Learning or

Artificial Intelligence. Artificial Intelligence (AI),
Deep Learning, one of the subfields of Machine machine learning and deep learning are three terms
Learning and Statistical Learning has been advancing often used interchangeably to describe software that
in impressive levels in the past years. Deep learning behaves intelligently. However, it is useful to
(also known as deep structured learning or understand the key distinctions among them.
differential programming) is part of a broader family
of machine learning methods based on artificial Artificial intelligence (AI) as an intelligence exhibited
neural networks with representation learning. [16] by machines has been an effective approach to human
Deep learning (DL) and Machine learning (ML) are learning and reasoning. AI is imparting a cognitive
two concepts related to Artificial Intelligence(AI). ability to a machine. The benchmark for AI is the
The two combined to improve the future of AI. They human intelligence regarding reasoning, speech, and
are different in many terms which must consider for vision. This benchmark is far off in the future.
creating the AI which is even better than the human Artificial Intelligence, as the name, suggests that it is
brain. Machine learning the outlook to get Artificial the intelligence created by humans. It constructed as
Intelligence and the implementation of it is known as complex machines using computer properties and
Deep learning. It is the difference which everyone performing various actions just like we the humans.
must know at the time of using these terms. Modern These machines have senses similar to humans, or if
technologies like AI, machine learning, deep learning, we say that they show and sense more than humans,
data science have become the buzzwords that then we are not wrong. In a nutshell, it is
everybody talks about to meet the current market incorporated human intelligence into machines.
opportunities. There is a misconception related to the These technology is the future of humanity and
words as most people think these things are the same making their lives better than before. The

IJSRSET207279 | Accepted : 02 April 2020 | Published : 09 April 2020 | March-April-2020 [ 7 (2) : 394-398]
394
 Mya Sandar Kyin et al Int J Sci Res Sci Eng Technol. March-April-2020; 7 (2) : 394-398

functionality of these technologies is similar to II. METHODS AND MATERIAL

humans, so they preferred as the best solution for the
tasks which can't be performed by us. If we try to Deep Learning (DL)
define this term, then there are no such words or Deep learning is a computer software that mimics the
definition which suits its properties. We can say that network of neurons in a brain. It is a subset of
it is a computer operated the machine which can machine learning and is called deep learning because
function similar to the human brain. [12] it makes use of deep neural networks. The machine
uses different layers to learn from the data. The depth
Machine learning is the best tool so far to analyze, of the model is represented by the number of layers in
understand and identify a pattern in the data. One of the model. Deep learning is the new state of the art in
the main ideas behind machine learning is that the term of AI. In deep learning, the learning phase is
computer can be trained to automate tasks that would done through a neural network. A neural network is
be exhaustive or impossible for a human being. The an architecture where the layers are stacked on top of
clear breach from the traditional analysis is that each other. Deep learning is a subfield of Machine
machine learning can take decisions with minimal Learning, and Machine Learning is a subfield of
human intervention. Machine learning uses data to Artificial Intelligence. Deep Learning is a sub-branch
feed an algorithm that can understand the of Machine Learning in that it also has a set of
relationship between the input and the output. When learning algorithms that can train on and learn from
the machine finished learning, it can predict the value data, and more specifically DL is powered by neural
or the class of new data point. Machine learning is a networks. Deep learning is based on a set of
part of Artificial Intelligence. The machines can learn. algorithms that attempt to model high-level
The robots learn themselves from the data provided abstractions in data by using multiple processing
to them. It makes more like to be a technique which layers with complex structures, or otherwise
makes us realize the presence of Artificial Intelligence. composed of multiple non-linear transformations [1].
This technique uses algorithms to get data, learn, and The deep learning methodology applies nonlinear
then analyze the data. The results came in the form of transformations and model abstractions of high level
predictions. We may have noticed when getting in large databases. Furthermore, the superior and
recommendation on shopping sites, Google, or beneficial of the deep learning methodology and its
Facebook. We get suggestions according to your hierarchy in layers and nonlinear operations are
interests. It is done with machine learning algorithms presented and compared with the more conventional
which are developed in the way to analyzing the algorithms in the common applications. Studies in
recent searches, history, and other information. This this area attempts to make better representations and
technique also influences the marketing and banking create models to learn these representations from
sectors. large-scale unlabeled data. Learn multiple levels of
representations that correspond to different levels of
Machine learning came directly from minds of the abstraction; the levels form a hierarchy of concepts
early AI crowd, and the algorithmic approaches over [1].
the years included decision tree learning, inductive
logic programming. clustering, reinforcement Deep learning exploits this idea of hierarchical
learning, and Bayesian networks among others. As we explanatory factors where higher level, more abstract
know, none achieved the ultimate goal of General AI, concepts are learned from the lower level ones. These
and even Narrow AI was mostly out of reach with architectures are often constructed with a greedy
early machine learning approaches. [13] layer-by-layer method. Deep learning helps to

International Journal of Scientific Research in Science, Engineering and Technology | www.ijsrset.com | Vol 7 | Issue 2
395
 Mya Sandar Kyin et al Int J Sci Res Sci Eng Technol. March-April-2020; 7 (2) : 394-398

disentangle these abstractions and pick out which multiple nonlinear layers (CAP > 2) and [4] considers
features are useful for learning [3]. For supervised CAP > 10 to be very deep learning.
learning tasks, deep learning methods obviate feature
engineering, by translating the data into compact Deep Learning Application
intermediate representations akin to principal Here have been several studies demonstrating the
components, and derive layered structures which effectiveness of deep learning methods in a variety of
remove redundancy in representation [1]. Many deep application domains. The reality of data proliferation
learning algorithms are applied to unsupervised and abundance of multimodal sensory information is
learning tasks. This is an important benefit because admittedly a challenge and a recurring theme in
unlabeled data is usually more abundant than labeled many military as well as civilian applications, such as
data. An example of a deep structure that can be sophisticated surveillance systems. Deep learning
trained in an unsupervised manner is a deep belief implies an abstract layer analysis and hierarchical
network [3]. methods. However, it can be utilized in numerous
real life applications. As an example, within digital
Deep Learning algorithm image processing; grayscale image coloring from a
Deep learning algorithms are based on distributed picture used to be done manually by users who had to
representations. The underlying assumption behind choose each color based on their own judgment.
distributed representations is that observed data is Applying a deep learning algorithm, coloring can be
generated by the interactions of factors organized in performed automatically by a computer [14. The
layers. Deep learning adds the assumption that these applications of deep learning are categorized in pure
layers of factors correspond to levels of abstraction or digital image processing, medicine and biometrics.
composition. Varying numbers of layers and layer
sizes can be used to provide different amounts of III. RESULTS AND DISCUSSION
abstraction [3]. Deep learning algorithms are
contrasted with shallow learning algorithms by the Practical Examples of Deep Learning
number of parameterized transformations a signal Deep learning allows machines to solve complex
encounters as it propagates from the input layer to the problems even when using a data set that is very
output layer, where a parameterized transformation is diverse, unstructured and inter-connected. The
a processing unit that has trainable parameters, such deeper learning algorithms learn, the better they
as weights and thresholds [4]. A chain of perform. Now that we’re in a time when machines
transformations from input to output is a credit can learn to solve complex problems without human
assignment path (CAP). CAPs describe potentially intervention, what exactly are the problems they are
causal connections between input and output and tackling? Here are just a few of the tasks that deep
may vary in length. For a feedforward neural network, learning supports today and the list will just continue
the depth of the CAPs, and thus the depth of the to grow as the algorithms continue to learn via the
network, is the number of hidden layers plus one (the infusion of data. The virtual assistants of online
output layer is also parameterized). For recurrent service providers use deep learning to help
neural networks, in which a signal may propagate understand your speech and the language humans use
through a layer more than once, the CAP is when they interact with them. Translations, in a
potentially unlimited in length. There is no similar way, deep learning algorithms can
universally agreed upon threshold of depth dividing automatically translate between languages. This can
shallow learning from deep learning, but most be powerful for travelers, business people and those in
researchers in the field agree that deep learning has government. Vision for driverless delivery trucks,

International Journal of Scientific Research in Science, Engineering and Technology | www.ijsrset.com | Vol 7 | Issue 2
396
 Mya Sandar Kyin et al Int J Sci Res Sci Eng Technol. March-April-2020; 7 (2) : 394-398

drones and autonomous cars is the way an being explicitly told to do so. This ability means that
autonomous vehicle understands the realities of the data scientists can sometimes save months of work.
road and how to respond to them whether it’s a stop Besides, the neural networks that a deep learning
sign, a ball in the street or another vehicle is through algorithm is made of can uncover new, more complex
deep learning algorithms. The more data the features that human can miss. The following are the
algorithms receive, the better they are able to act benefits of Deep Learning, features are automatically
human-like in their information processing knowing deduced and optimally tuned for desired outcome and
a stop sign covered with snow is still a stop sign. are not required to be extracted ahead of time, avoids
Chatbots and service bots that provide customer time consuming machine learning techniques,
service for a lot of companies are able to respond in robustness to natural variations in the data is
an intelligent and helpful way to an increasing automatically learned same as neural network based
amount of auditory and text questions thanks to deep approach can be applied to many different
learning. Image colorization, the transforming black- applications and data types and the massive parallel
and-white images into color was formerly a task done computations can be performed using GPUs and are
meticulously by human hand. Today, deep learning scalable for large volumes of data. Moreover, it
algorithms are able to use the context and objects in delivers better performance results when amount of
the images to color them to basically recreate the data is huge. [7]
black-and-white image in color. The results are
impressive and accurate. Facial recognition of deep Drawbacks of Deep Learning
learning is being used for facial recognition not only In order to solve a problem, deep learning enables
for security purposes but for tagging people on machines to mirror the human brain by making use of
Facebook posts and we might be able to pay for items artificial neural networks. These networks are known
in a store just by using our faces in the near future. to run a variety of applications such as speech
The challenges for deep-learning algorithms for facial recognition devices like Siri and Neuro-Linguistic
recognition is knowing it’s the same person even Programming. Deep learning has hence been
when they have changed hairstyles, grown or shaved recognized as one of the major research areas required
off a beard or if the image taken is poor due to bad to advance AI. Although the importance of deep
lighting or an obstruction. learning is increasing and several advances in its
research are touching great heights, there are a few
Medicine and pharmaceuticals, disease and tumor downsides or challenges that have to be tackled to
diagnoses to personalized medicines created develop it. To exemplify, for a speech recognition
specifically for an individual’s genome, deep learning program, data formulating multiple dialects,
in the medical field has the attention of many of the demographics and time scales is required to obtain
largest pharmaceutical and medical companies. [15] desired results [8]. Moving on, though deep learning
models are very efficient and are able to formulate an
Benefits of Deep Learning adequate solution to a particular problem once trained
The deep learning architecture is flexible to be with data, they are unable to do so for a similar
adapted to new problems in the future One of deep problem and require retraining [10]. To elaborate,
learning’s main advantages over other machine these neural network architectures are highly
learning algorithms is its capacity to execute feature specialized to a specific domain and reassessment is
engineering on its own. A deep learning algorithm needed to solve issues that do not pertain to that
will scan the data to search for features that correlate identical domain. Another challenge of deep learning
and combine them to enable faster learning without is it requires large amounts of processing power. This

International Journal of Scientific Research in Science, Engineering and Technology | www.ijsrset.com | Vol 7 | Issue 2
397
 Mya Sandar Kyin et al Int J Sci Res Sci Eng Technol. March-April-2020; 7 (2) : 394-398

high-performance hardware is mostly the multi-core [4]. J. Schmidhuber, "Deep learning in neural
high performing graphics processing unit or a similar networks: An overview," Neural Networks, vol.
processing system [8]. These processing units require 61, pp. 85-117, 2015.
and consume a lot of power and are therefore a costly [5]. Y. LeCun, Y. Bengio, and G. Hinton, "Deep
affair. Computing Hardware: As you mentioned, the learning," Nature, vol. 521, pp. 436-444, 2015.
GCPUs and alike processors have limitations. Our [6]. I. Arel, D. C. Rose, and T. P. Karnowski, "Deep
speaker from IBM in class 3 had touched the subject machine learning-a new frontier in artificial
intelligence research research frontier],"
of quantum computation. That solution looks very
Computational Intelligence Magazine, IEEE, vol.
promising for reducing computation time and
5, pp. 13-18, 2010.
complexity. That is a better future to reduce
[7]. https://becominghuman.ai/what-is-deep-learning-
computation complexity needed by DL. [9][11]
and-its-advantages-16b74bc541a1
[8]. https://hackernoon.com/challenges-in-deep-
IV.CONCLUSION
learning-57bbf6e73bb
[9]. https://www.theverge.com/2016/10/10/13224930/
Deep learning is powerful technique in current ai-deep-learning-limitations-drawbacks
generation and will have many more successes in the [10]. https://bdtechtalks.com/2018/02/27/limits-
near future because it requires very little engineering challenges-deep-learning-gary-marcus/
by hand, so it can easily take advantage of increases in [11]. https://becominghuman.ai/what-i-learned-from-
the amount of available computation and data. andrew-ngs-deep-learning-specialization-
Artificial intelligence in recent times, coupled with ccf94fea2a0
the recognition that deep learning is evolving as one [12]. https://www.investopedia.com/terms/a/artificial-
of its most powerful techniques. Though this paper intelligence-ai.asp
has tried to present an overview of Artificial [13]. Mitchell, T. M. Machine Learning. New York:
Intelligence, Machine Learning, deep learning and McGraw-Hill, 1997
how its advantages and drawback of regarding deep [14]. M. Osadchy, Y. L. Cun, and M. L. Miller,
learning algorithms and its applications, there still "Synergistic face detection and pose estimation
remains a great deal of work to be done in improving with energy-based models," The Journal of
the learning process. Machine Learning Research, vol. 8, pp. 1197-1215,
2007
V. REFERENCES [15]. https://www.forbes.com/sites/bernardmarr/2018/1
0/01/what-is-deep-learning-ai-a-simple-guide-
with-8-practical-examples/#a848b5c8d4ba.
[1]. D. Li and D. Yu, "Deep Learning: Methods and
[16]. https://en.wikipedia.org/wiki/Deep_learning
Applications," Foundations and Trends in Signal
Processing, Now Publishers, 2014.
Cite this article as :
[2]. Y. Bengio, "Learning deep architectures for AI,"
Mya Sandar Kyin, Zaw Lin Oo, Khin Mar Cho, "An
Foundations and trends® in Machine Learning,
Overview Studying of Deep Learning", International
vol. 2, pp. 1-127, 2009.
Journal of Scientific Research in Science, Engineering
[3]. Y. Bengio, A. Courville, and P. Vincent,
and Technology (IJSRSET), Online ISSN : 2394-4099,
"Representation learning: A review and new
Print ISSN : 2395-1990, Volume 7 Issue 2, pp. 394-398,
perspectives," Pattern Analysis and Machine
March-April 2020. Available at doi :
Intelligence, IEEE Transactions on, vol. 35, pp.
https://doi.org/10.32628/IJSRSET207279
1798-1828, 2013.
Journal URL : http://ijsrset.com/IJSRSET207279

International Journal of Scientific Research in Science, Engineering and Technology | www.ijsrset.com | Vol 7 | Issue 2
398