International Journal of Innovative Technology and Exploring Engineering (IJITEE)

ISSN: 2278-3075, Volume-8 Issue-12, October 2019

Machine Learning and Deep Learning

Ayushi Chahal, Preeti Gulia


Abstract: Now-a-days artificial intelligence has become an
asset for engineering and experimental studies, just like statistics
and calculus. Data science is a growing field for researchers and
artificial intelligence, machine learning and deep learning are
roots of it. This paper describes the relation between these roots of
data science. There is a need of machine learning if any kind of
analysis is to be performed. This study describes machine learning
from the scratch. It also focuses on Deep Learning. Deep learning
can also be known as new trend of machine learning. This paper
gives a light on basic architecture of Deep learning. A
comparative study of machine learning and deep learning is also
given in the paper and allows researcher to have a broad view on Fig. 1. AI, machine learning and deep learning
these techniques so that they can understand which one will be paradigm
preferable solution for a particular problem.
II. MACHINE LEARNING
Keywords : Machine Learning, Deep learning, Artificial
Intelligence, shallow learning. Machine learning is based on the idea that system can learn
from data, identify the patterns and make decision with
I. INTRODUCTION minimum human intervention [2]. This is the scientific study
of algorithms and statistical models with the help of which
In the era of data sciences, artificial intelligence is trying computer system perform a specific task without using
instruction, inference and patterns. Machine learning
to provide human kind intelligence to the computer and for
algorithms build mathematical model based on sample data
this machine learning and deep learning are the technologies
and then make the decision.
which are helping artificial intelligence to do it. Machine
Learning is the branch or subset of artificial intelligence that A. Machine learning procedure
train the machines how to learn. Deep learning is confined Machine learning incorporates four steps, given below
version of machine learning. It helps to raise the high (shown in the figure 2):
standards of learning environment. Machine learning and - First, feature extraction
deep learning both plays vital role in upgrading the computer - Second, selection of corresponding machine learning
systems to be an expert systems that can take decisions and algorithm
make predictions without a human intervention. - Third, training and evaluation the data model’s efficiency
Artificial intelligence is a field which helps computer - Four, using trained model for prediction
system to be intelligent and take decisions. Machine learning
B. Requirements to Create Good Machine Learning
helps to implement Artificial Intelligence on the system and Systems:
deep learning helps to achieve machine learning goals on the
 Data preparation capabilities
system more systematically. Figure 1 shows it pictorially.
 Basic and Advanced algorithms
This paper is divided into two parts. In section II, will
 Scalability
explain machine learning, its procedures, its applications etc.  Various processes i.e. Automation and Iterative
In section III, Different approaches of machine learning are  Ensemble modeling
discussed such as deep learning and shallow learning. It
describes each one’s different methods and different
algorithms used by them. In section IV, a comparative study
between deep learning and other conventional methods of
machine learning.

Revised Manuscript Received on October 05, 2019.

* Correspondence Author
Ayushi Chahal*, Department of Computer Science and Applications,
Maharishi Dayanand University, Rohtak, India. Email:
[email protected]
Preeti Gulia, Department of Computer Science and Applications,
Maharishi Dayanand University, Rohtak, India. Email:
[email protected]

Retrieval Number: L35501081219/2019©BEIESP Published By:

DOI: 10.35940/ijitee.L3550.1081219 Blue Eyes Intelligence Engineering
4910 & Sciences Publication
 Machine Learning and Deep Learning

Fig. 2. Machine learning model

 Government sector: Government agencies use machine
C. Relationship with Other Fields:
learning to mine the data for insight where agencies like
Machine learning is considered as the subset of artificial public safety and utilities etc. have multiple sources of data.
intelligence. In earlier days of AI as academic discipline, Sensor data analysis increases the efficiency and save money.
researchers were interested in having machine learn. They Machine learning can also be used for security purpose i.e.
attempted to solve the problem with various symbolic help to detect fraud and to minimize the identity theft.
methods as well as connectionist approach where neural  Retail sector: In retail sector, machine learning is used
network, pattern recognition are used. In 1990s, Machine to analyze the buying history of customers. Retailers rely on
learning is reorganized as a separate field. It shifted focus machine learning to capture data, analyze and use it to
from symbolic approach to the methods and models of personalize the shopping experience. It is also helpful to
statistics and probability theory [4]. implement the marketing campaign, optimizing price, and for
 Relation to data mining: Both of these employ same customer insights.
methods often and overlap with each other. But machine  Transportation: Machine learning is used to make
learning focuses on prediction based on known properties routes more efficient and to predict the problems to increase
while data mining focuses on the discovery of unknown profitability. It can be done after analyzing the data to identify
properties. Data mining uses machine learning methods, patterns and trends. Data analysis and modeling aspects are
key factors to delivery companies and transportation
machine learning also employs data mining methods; but with
organizations.
different goals or to improve the learner accuracy.
 Oil and gas: In this sector, machine learning is used to
 Relation to optimization: Machine learning is also
find new energy source and to analyze minerals in ground. It
intimated with optimization. Learning problems are is also used to predict refinery sensor failure. Streamlining
formulated as minimization of loss function. Loss functions oil distribution makes it more efficient and economic.
show the discrepancy between prediction of model and actual E. Processes and Techniques associated with
problem. machine learning:
 Relation to statistics: It is also closely related with A number of processes, techniques and methods can be
statistics. The ideas of machine learning have had a applied to enhance the performance of machine learning and
relationship with statistics from methodological principles to these are as follows:
theoretical tools such as the modeling paradigm.  Feature learning
 Sparse dictionary learning
D. Who’s Using Machine Learning?  Anomaly detection
As the industries grow, large volumes of data have been  Decision tree
recognized. For handling that data, machine learning  Association rules
technology is required. With the machine learning, F. Applications of Machine learning:
organizations are able to work more efficiently. Machine There are many applications of machine learning such as:
learning is used in following areas:  Adaptive websites
 Financial services: In financial services, machine  Bioinformatics
learning technology is used to identify the important insight in  Brain-machine interface
data and to prevent fraud. The insights help to identify  Computer vision
investment opportunities or help investors to know when to  Data quality
trade. Data mining concepts also identify high risk profiles of  DNA sequence classification handwriting recognition
clients or to pinpoint warning signs of fraud.  Machine learning control
 Health Care: This is the major area in which wearable  User behavior analytics etc…
devices and sensors are used to assess patient’s health in real
time. Machine learning also helps medical experts to analyze
the data to identify trends. This may lead to improve
diagnoses and treatment.

Retrieval Number: L35501081219/2019©BEIESP Published By:

DOI: 10.35940/ijitee.L3550.1081219 Blue Eyes Intelligence Engineering
4911 & Sciences Publication
 International Journal of Innovative Technology and Exploring Engineering (IJITEE)
ISSN: 2278-3075, Volume-8 Issue-12, October 2019

III. MACHINE LEARNING APPROACHES  Semi-supervised learning: In some cases, input may
be only partially available, or restricted to special feedback.
Basically, Machine learning methods are broadly
categorized in two categories i.e. Shallow learning and deep At that time, these algorithms are used. These are used to
learning [16]. Shallow learning basically uses neural networks develop mathematical model from incomplete training data,
with single layers or SVMs (Support Vector Machines) while where a portion of the sample input doesn’t consist
deep learning uses neural network with more than one hidden labels. This learning is useful when cost of labeling is too
layers. As shown in figure 3: high to allow for fully labeled training process.
 Reinforcement learning: This is the area of learning
concerned with how software agents take actions in an
environment to maximize the cumulative reward. In this
type of learning, a feedback is to be given in the form of
positive or negative reinforcement in a dynamic
environment. These are commonly used in autonomous
vehicle or in learning to play game against human opponent
[3]. Q-learning is an example of reinforcement learning.
 Active learning: Desired outputs are accessed for a
limited set of inputs. In this learning, the inputs are based on
Fig. 3. Different approach of machine learning budget, and optimize the choice of inputs for which output
A. Shallow Learning will be acquired.
 Meta learning: Here, algorithms learn their own
Shallow learning is broadly divided into two categories:
inductive bias based on previous experiences. Some
Supervised and Unsupervised Learning. But there are also
other methods of machine learning. Overview of popular examples of meta learning are Bagging, Boosting, Random
methods is as follows: Forest.
 Supervised learning: In supervised learning,
algorithm builds a mathematical model from a set of data
that contains both the input and desired outputs.(wiki)
These algorithms are trained using labeled examples i.e.
input and desired outputs are known. In this learning,
algorithm receives a set of inputs along with corresponding
correct outputs. Algorithm learns by comparing its actual
output with correct outputs to find out errors. Then, model
is modified accordingly. Classification, regression,
prediction and gradient boosting are the example of
supervised learning which use pattern to predict the values.
This learning is commonly used in those applications where
historical data predicts future events. Classification and Fig. 5. Unsupervised Learning
regression are the tasks that are performed by supervised
learning. Some examples of supervised machine learning B. Deep Learning
are Nearest neighbor, Naïve Bayes, Decision Tree, Deep learning is a set of algorithms of machine learning
Regression Tree etc. Figure 4 gives the pictorial view of which uses multiple layers that corresponds to different level
different method of supervised learning. of abstraction to each level. . It consists of input layer, output
 Unsupervised learning: In unsupervised learning, a layer and several hidden layer. It is used for voice synthesis,
mathematical model is to be built from a set of data which image processing, handwriting recognition, object detection,
contains only inputs. Desired output labels are not present prediction analytics and decision making. [10] Deep learning
in this type of learning. Unsupervised learning is used can be broadly classified into three types (figure 6):
against that data which doesn’t consists historical label.
K-means, Association Rules are an example of such
algorithms. Figure 5 describes different methods of
unsupervised learning.

Fig. 6. Deep learning

Fig. 4. Supervised Learning

Retrieval Number: L35501081219/2019©BEIESP Published By:

DOI: 10.35940/ijitee.L3550.1081219 Blue Eyes Intelligence Engineering
4912 & Sciences Publication
 Machine Learning and Deep Learning

 Generative models: Generative models are used for C. Deep Learning architecture
unsupervised learning. It includes algorithms like Deep Deep Learning consists of supervised or unsupervised learning
Belief Network (DBN), Deep auto-encoders, Deep techniques based on many layers of artificial neural networks
Boltzmann (DBM). that are able to learn hierarchical representations in deep
 Discriminative models: Discriminative models usually architectures. [11] It is extended version of artificial neural
provide supervised learning approaches. It involves network. Deep Learning architectures consist of multiple
Convolution Neural Network (CNN), Deep Stacking processing layers. Each layer is able to produce non-linear
Network (DSN). responses based on the data from its input layer.
 Hybrid models: Hybrid models incorporate the benefits of The functionality of Deep Learning is imitated from the
both discriminative and generative models. Deep Neural mechanisms of human brain and neurons for processing of
network (DNN) is an example of hybrid models.

Fig. 8. Deep Neural network

signals. Deep Learning architectures have gained more
attention in recent years compared to the other traditional IV. DEEP LEARNING COMPARISON WITH
machine learning approaches. Such approaches are considered CONVENTIONAL MACHINE LEARNING
as being shallow-structured learning architectures versions TECHNIQUES
(i.e., a limited subset) of Deep Learning.
Deep learning is a new era of machine learning. Deep
A Deep Neural Network consists of an input layer,
learning includes both supervised and unsupervised learning
severalhidden layers, and an output layer. Each layer includes
paradigm of machine learning. Machine learning and deep
severalunits called neurons. These neurons are also called as
learning helps in providing intelligence to the system that can
artificial neurons. A neuron receives several inputs, performs
make prediction for future using past data.[12]
a weighted summation over its inputs, then the resulting sum
 Conventional machine learning algorithms can’t learn
goes through an activation function to produce an output.
directly from the raw data. They need careful engineering
Each neuron has a vector of weights associated to its input
to carefully extract features from raw data and highly
size as well as a bias that should be optimized during the
classified domain expertise, which are further used to in
training process. Figure 7 below shows the structure of
internal representations to identify these feature’s
neuron.
patterns. In Deep Learning, first step of machine learning
procedure is not present. This step is automated in deep
learning. Deep Learning can extract new features
automatically from raw data. Figure 9 shows this point
clearly [13].

Fig. 7. Structure of a neuron

When these artificial neurons are assigned sequentially which
makes a chain as one neurons output becomes input of next
neuron, and this process goes on over and over which makes a
Artificial Neural Network. Deep Learning Neural Networks
consists of more than one hidden layer as shown below in
figure 8. Fig. 9. Feature extraction is automated in deep
learning

Retrieval Number: L35501081219/2019©BEIESP Published By:

DOI: 10.35940/ijitee.L3550.1081219 Blue Eyes Intelligence Engineering
4913 & Sciences Publication
 International Journal of Innovative Technology and Exploring Engineering (IJITEE)
ISSN: 2278-3075, Volume-8 Issue-12, October 2019

 Deep learning algorithms work more accurately on large 6. Why Machine Learning Models Often Fail to Learn: QuickTake
Q&A". Bloomberg.com. 2016-11-10. Retrieved 2017-04-10.
Data set as compared to conventional machine learning 7. "The First Wave of Corporate AI Is Doomed to Fail". Harvard Business
algorithms. While machine learning algorithms Review. 2017-04-18. Retrieved 2018-08-20.
outperforms deep learning in case of small or medium size 8. "Why the A.I. euphoria is doomed to fail". VentureBeat. 2016-09-18.
datasets. [14] Retrieved 2018-08-20.
9. "9 Reasons why your machine learning project will
 Deep learning algorithms take less time to infer a problem fail". www.kdnuggets.com. Retrieved 2018-08-20.
as compared to conventional machine learning algorithms 10. P. Kuang, W.-N. Cao, and Q. Wu, “Preview on structures and
 Deep learning performs a high amount of matrix multiple algorithms of deep learning,” in 2014 11th International Computer
Conference on Wavelet Actiev Media Technology and Information
hence it needs powerful engine preferably GPU (Graphical Processing(ICCWAMTIP), Chengdu, China, 2014, pp. 176–179.
Processing Units) or specially designed TPU (Tensor 11. L. Deng, “Deep Learning: Methods and Applications,” Found.
Processing Units) while other conventional machine Trends® Signal Process., vol. 7, no. 3–4, pp. 197–387, 2014.
12. J. L. Berral-García, “When and How to Apply Statistics, Machine
learning algorithms can work on low end machines. Learning and Deep Learning Techniques,” in 2018 20th International
 Deep learning algorithms are difficult to impossible to Conference on Transparent Optical Networks (ICTON), 2018, pp.
interpret. Some of the machine learning algorithms like 1–4.
13. J. Latif, C. Xiao, A. Imran, and S. Tu, “Medical Imaging using
(logistics, decision tree) can be interpreted easily while
Machine Learning and Deep Learning Algorithms: A Review,” in 2019
some (like SVM) are almost impossible to interpret. [15] 2nd International Conference on Computing, Mathematics and
 Training time for data to create the model is more in deep Engineering Technologies (iCoMET), 2019, pp. 1–5.
learning as compared to other machine learning algorithms. 14. N. G. Paterakis, E. Mocanu, M. Gibescu, B. Stappers, and W. van Alst,
“Deep learning versus traditional machine learning methods for
aggregated energy demand prediction,” in 2017 IEEE PES Innovative
V. CONCLUSION Smart Grid Technologies Conference Europe (ISGT-Europe), 2017,
pp. 1–6.
This article examined the concepts of machine learning. 15. M. Sewak, S. K. Sahay, and H. Rathore, “Comparison of Deep
Machine learning has gained a lot of attention of researchers Learning and the Classical Machine Learning Algorithm for the
nowadays due to its distinct features. Firstly, the article Malware Detection,” in 2018 19th IEEE/ACIS International
Conference on Software Engineering, Artificial Intelligence,
specified the points to make a good machine learning system. Networking and Parallel/Distributed Computing (SNPD), 2018, pp.
Followed by this, the usage and applications of machine 293–296.
learning have been discussed in this article. However the road 16. A. Jabeen, N. Ahmad, and K. Raza, “Machine Learning-based
of machine learning is not as simple as it looks to be. There state-of-the-art methods for the classification of RNA-Seq data,”
Bioinformatics, preprint, Mar. 2017.
are some challenges in this area to get the expected results
such as lack of suitable data, data bias, and lack of resources,
privacy problems and evaluation problems. This paper crates AUTHORS PROFILE
a broad view for a researcher for machine learning by
categorizing it into two parts, namely: shallow learning and Ms. Ayushi Chahal has completed her M.Tech
from GJUS&T University. She is currently
deep learning. Supervised and unsupervised machine learning pursuing Ph.D. in Computer Science at Department
concepts are supposed to be in the category of shallow of Computer Science & Applications,
learning as these techniques use less number of hidden layers M.D.University, Rohtak. Her main research area
or SVMs. While deep learning is considered as a different includes network security, Internet of Things (IoT),
Machine Learning.
category, because of its deep layered architecture discussed in
the article.
Deep learning is a growing field in a sector of predictive Dr. Preeti Gulia is currently working as
analytics. This paper provides a comparative study of Assistant Professor at Department of Computer
Science & Applications, M.D.University, Rohtak,
conventional methods of machine learning and deep learning
India. She is serving the Department since 2009.
which helps new researchers to choose which technique She earned her doctoral degree in 2013. She has
would be right to apply in a particular environment. Such as, if published more than 65 research papers and articles
one is working on small training data set then he must use in journal and conferences of National/
machine learning algorithms rather than deep learning while, International repute including ACM, Scopus. Her
area of research includes Data Mining, Big Data,
if dataset needed to choose the features then one must use Machine Learning, Deep Learning, IoT, Software Engineering. She is an
machine learning technique because in case of deep learning active professional member of IAENG, CSI and ACM. She is also serving as
this feature selection procedure is automated researcher do Editorial Board Member Active Reviewer of International/ National
not have to bother about it. This paper creates base for the Journals. She has guided one research scholar as well as guiding four Ph.D.
research scholars from various research areas..
researcher who wants to pursue research in field of artificial
intelligence or predictive analytics.

REFERENCES
1. https://en.wikipedia.org/wiki/Machine_learning
2. https://www.sas.com › SAS Insights › Analytics Insights
3. Bishop, C. M. (2006), Pattern Recognition and Machine Learning,
Springer, ISBN 978-0-387-31073-2
4. 11. Langley, Pat (2011). "The changing science of machine
learning". Machine Learning. 82 (3): 275–
279. Doi:10.1007/s10994-011-5242-y
5. Alpaydin, Ethem (2010). Introduction to Machine Learning. MIT
Press. p. 9. ISBN 978-0-262-01243-0

Retrieval Number: L35501081219/2019©BEIESP Published By:

DOI: 10.35940/ijitee.L3550.1081219 Blue Eyes Intelligence Engineering
4914 & Sciences Publication