International Journal of Scientific & Engineering Research Volume 10, Issue 9, September-2019

ISSN 2229-5518 1536

Machine Learning and Deep Learning

Revolutionize Artificial Intelligence

Adil KORCHI, Fayçal MESSAOUDI, Lahcen Oughdir

Abstract— The concept of Artificial Intelligence (AI) is present in many aspects of our professional life. From automatic translation to
predictive maintenance, to infused intelligence in the applications you use every day, it's not about the future, but about the company
present. The fields of application and potential uses of Artificial Intelligence are more and more diverse: understanding of natural language,
visual recognition, robotics, autonomous system, Machine Learning.

In this manuscript, we will show what is the Machine Learning concept and the Deep Learning as well as their position in artificial
intelligence, their strengths and their flaws. We will present some algorithms that the Learning machine uses. We will also d iscuss the
statistics used in these algorithms to adapt the links between the Artificial Neural Networks to strengthen or destroy the links in order to
have a good approximation of the input data.

Index Terms— Artificial intelligence - Machine Learning - Deep Learning - Algorithms - Statistics - Artificial Neural Networks - Cloud
Computing - Big Data

——————————  ——————————

IJSER
1 INTRODUCTION

T his manuscript explains the power of artificial intelligence,

machine learning and deep learning as well as the rela-
tionship between them. It also discusses statistical curves that
2.2 Machine Learning (ML)
It is a concept that is more and more talked about in the world
approximate data and make it easy to generalize and cre- ate of computing, and that relates to the field of artificial intelli-
algorithms that can receive a lot of data to make a decision gence. Still called "statistical learning", this term refers to a
and create a model to automate tasks according to situations. process of development, analysis and implementation leading
It also talks about Deep Learning and explains how it is in- to the establishment of systematic processes. To put it simply,
spired by our brain (with networks of neurons) to push the it is a kind of program that allows a computer or a machine an
analysis further and know how to extract the data itself. automated learning, so that it can perform a number of very
complex operations. The aim is to make the machine or com-
puter capable of providing solutions to complicated problems
by processing an astronomical amount of information. This
2 DEFINITIONS offers an opportunity to analyze and highlight the correlations
2.1 Artificial Intelligence that exist between two or more given situations, and to predict
Artificial intelligence can be defined as "the set of theories and their different implications.
techniques used to create machines capable of simulating in-
telligence (Krishnamoorthy, 2018). Either computers or pro-
2.3 Deep Learning (DL)
grams capable of performance usually associated with human
intelligence, and amplified by technology, namely: Deep Learning is a subdomain of Machine Learning. It is
1. Ability to reason a form of artificial intelligence, derived from Machine
2. Ability to process large amounts of data Learning. To understand what Deep Learning is, it is im-
3. Ability to discern patterns and patterns undetectable portant to understand what machine learning is. Deep
by a human learning usually has two phases. The first is to estimate a
4. Ability to understand and analyze these models model from data, called observations, that are available
5. Ability to interact with humans and in finite numbers, during the design phase of the sys-
6. Faculty to learn gradually tem (Goodfellow, 2016). Model estimation involves solving
7. And continually improve its performance a practical task, such as translating a speech, estimating a
probability density, recognizing the presence of a cat in a
"Artificial intelligence" thus covers a vast subject, in per- photograph, or participating in the driving of an autono-
petual mutation. And the dazzling progress since 1950, the mous vehicle. This so-called "learning" or "training" phase
founding year of AI. is generally performed prior to the practical use of the

IJSER © 2019
http://www.ijser.org
 International Journal of Scientific & Engineering Research Volume 10, Issue 9, September-2019
ISSN 2229-5518 1537

model. The second phase corresponds to the setting in learn.

production: the model being determined, new data can Machine Learning, or machine learning, is able to reproduce
then be submitted in order to obtain the result correspond- behavior thanks to algorithms, themselves powered by a large
ing to the desired task. In practice, some systems can con- amount of data. Faced with many situations, the algorithm
tinue their learning once in production, provided they have learns the decision to adopt and creates a model. The machine
a way to get a return on the quality of the results produced. can automate tasks according to the situations.
To better understand this principle, we give the following ex-
ample:
Before 1980, programs were made to calculate the price of
apartments according to their area, for example, "if the area is
less than 30m ², the price is worth € 65 000, if it is between 30m
² and 50m ², the price is worth € 85,000, etc ... ", or maybe"
price = area * € 3,500 "
These approximations are not satisfactory, and it would suf-
fice to note the price of the apartment knowing its true area to
estimate the price of a new flat of non-referenced. By adopting
this reasoning, we give birth to machine learning. So we can
Fig. 1. Place of Deep Learning in the world of computing create statistical curves that approximate the data and make it
easy to generalize and create algorithms that can receive a lot
of data.
Deep learning has yielded impressive results in many Machine learning is a broad field, which includes many algo-
different fields: rithms. Among the most famous, we find:

• Regressions (linear, multivariate, polynomial, regularized,

1. Recognition of image, text, voice, face ... logistic ...): these are curves that approximate the data (see
2. Segmentation in the medical field, understanding a diagram above)

IJSER
scene, a text ...
3. Generation of image, text, voice, artwork, human
face ... Autonomous car, autonomous robot ...
4. Road monitoring, pedestrian ...
5. The victory in the GO game against the world
champion of this discipline.

We could still mention many more original examples

than others, but what must be remembered is that deep
learning makes it possible to teach a computer a precise
task by observing a large number of examples.

3 SECTIONS MACHINE LEARNING AND DEEP LEARNING

ARE ARTIFICIAL INTELLIGENCE
3.1 Machine Learning
Fig. 2. Data approximation curves
Artificial Intelligence as we know it is a weak Artificial Intelli-
gence, as opposed to Strong Artificial Intelligence, which does
not exist yet (Hamming, 2017). Today, machines are able to
reproduce human behavior, but without awareness. Later,  The Naïve Bayes algorithm: the algorithm gives the
their abilities could grow to the point of becoming machines probability of the prediction, knowing the previous
with consciousness, sensitivity and spirit. events (El Kourdi, 2004). For example, what is the
If Machine Learning and Deep Learning are Artificial Intelli- most likely price of an 87.6m² apartment.
gences, the opposite is not true. For example, knowledge  Clustering: always thanks to mathematics, we will
graphs or rule engines are Artificial Intelligences but are not group the data in packets so that in each packet the
Machine Learning or Deep Learning. Deep Learning is a data are as close as possible to each other. It is used
branch of Machine Learning. especially for recommendations of movies "close"
Artificial Intelligence has evolved a lot thanks to the emer- movies you have already seen!
gence of Cloud Computing and Big Data (Beam, 2018), a low-  Decision trees: by answering a certain number of
cost computing power and accessibility to a large amount of questions and following the branches of the tree that
data. Thus, the machines are no longer programmed; they carry these answers, we arrive at a result (with a

IJSER © 2019
http://www.ijser.org
 International Journal of Scientific & Engineering Research Volume 10, Issue 9, September-2019
ISSN 2229-5518 1538

probability score) Thus was born the idea of Deep Learning towards the year 2010: to
 More sophisticated algorithms based on several statis- take inspiration from how our brain works (with networks of neu-
tical techniques: Random Forest (a forest of decision rons) to push the analysis further and know how to extract the data
trees that vote), Gradient Boosting, Support Vector itself!
Machine ... Deep learning is therefore based on what are called artificial neural
networks (deep), that is to say a set of neurons (they are small cal-
culators that perform a mathematical operation) that send them-
selves numbers in function of their links, up to output neurons
(Hertz, 2018). Thanks to this architecture, the Deep Learning is able
to recognize faces, to synthesize texts or to drive an autonomous
car!

Fig. 5. Architecture of an artificial neural network

Figure 3: Naive Bayes algorithm (multinomial model): Training and
testing.

IJSER
The next step is to use the statistics for the algorithm to adapt the
links between its neurons to reinforce or destroy them, to ensure
that at the output we have a good approximation of the input data.
3.2 Deep Learning For example, in a network that has learned to predict the price of
an apartment, if we give in input "30", then out we will have a
number very close to "65,000".

We can say that :

DL = ML + NEURONS NETWORKS
where DL extracts automatically relevant data information and
adapts its algorithm.

3.3 Examples of Deep Learning algorithms:

Artificial neural networks (ANN): these are the simplest and are
often used in addition because they sort the information well

 Convolutional Neural Networks (CNNs): specialized

Fig 4: Artificial Neural Networks (Deep Learning) in image processing, they apply filters to data to bring
out new information (for example, highlighting out-
lines in an image can help find where the face)
 Recurrent neural networks (RNN): the most well
Despite its power, Pure Machine Learning has many flaws. The
known are the LSTMs, which have the advantage of
first is that a human expert must sort the data to remove any un-
retaining information and reusing it soon after
necessary information and could clutter or penalize the system.
(Snoek, 2012). They are used for the analysis of text
For example, for the sale of an apartment, if you think that the
(NLP), since each word depends on the few previous
owner's age does not affect the price, there is no point in giving this
words (so that the grammar is correct)
information to the algorithm because if you give the system too
 More advanced versions, such as auto-encoders,
much, it could see relationships where there are none.
Boltzmann machines, self-organizing maps (SOM) ...
Then, the second (which follows from the first): how to recognize a
face? You could give the algorithm a lot of information about the
person (gap between the eyes, forehead height, etc ...), but it would
not be very adaptive or precise.
The figure below shows where machine learning and deep learn-

IJSER © 2019
http://www.ijser.org
 International Journal of Scientific & Engineering Research Volume 10, Issue 9, September-2019
ISSN 2229-5518 1539

ing are in artificial intelligence. ods for large-scale machine learning. Siam Review, 2018, vol. 60, no 2, p. 223-
311.
[4] BRADLEY, Andrew P. The use of the area under the ROC curve in the evalu-
ation of machine learning algorithms. Pattern recognition, 1997, vol. 30, no 7,
p. 1145-1159.
[5] DENG, Li, YU, Dong, et al. Deep learning: methods and applica-
tions. Foundations and Trends® in Signal Processing, 2014, vol. 7, no 3–4, p.
197-387.
[6] DIETTERICH, Thomas G. Ensemble methods in machine learning. In
: International workshop on multiple classifier systems. Springer, Berlin, Hei-
delberg, 2000. p. 1-15.
[7] DONG, Chao, LOY, Chen Change, HE, Kaiming, et al. Learning a deep con-
volutional network for image super-resolution. In : European conference on
computer vision. Springer, Cham, 2014. p. 184-199.
[8] EL KOURDI, Mohamed, BENSAID, Amine, et RACHIDI, Tajje-eddine. Au-
tomatic Arabic document categorization based on the Naïve Bayes algorithm.
In : proceedings of the Workshop on Computational Approaches to Arabic
Script-based Languages. Association for Computational Linguistics, 2004. p.
51-58.
[9] FERBER, Jacques et WEISS, Gerhard. Multi-agent systems: an introduction to
Fig. 6 : DL and ML used with neural networks using Artificial intelli- distributed artificial intelligence. Reading : Addison-Wesley, 1999.
gence [10] GOLDBERG, David E. et HOLLAND, John H. Genetic algorithms and ma-
chine learning. Machine learning, 1988, vol. 3, no 2, p. 95-99.
[11] GOODFELLOW, Ian, BENGIO, Yoshua, et COURVILLE, Aaron. Deep learn-
ing. MIT press, 2016.
[12] HAMMING, Richard W. et FEIGENBAUM, Edward A. Problem solving

IJSER
4 CONCLUSION
methods in artificial intelligence. 2017.
Machine learning and deep learning are very powerful tools [13] HERTZ, John A. Introduction to the theory of neural computation. CRC Press,
that make it possible to perform multiple actions such as clas- 2018.
sifying data, teaching a program from experiments, or creating [14] KRISHNAMOORTHY, C. S. et RAJEEV, S. Artificial intelligence and expert
an evolutionary program that is constantly improving. Thus, systems for engineers. CRC press, 2018.
even with a little provided sample and data influenced by the [15] LU, Huimin, LI, Yujie, CHEN, Min, et al. Brain intelligence: go beyond artifi-
subjectivity of the one whom measures them, these two tools cial intelligence. Mobile Networks and Applications, 2018, vol. 23, no 2, p. 368-
remain relatively precise despite some shortcomings. 375.
Nevertheless, the machine and the deep learning do not have [16] MÜLLER, Vincent C. et BOSTROM, Nick. Future progress in artificial intelli-
only qualities, they must be constantly adapted to the prob- gence: A survey of expert opinion. In : Fundamental issues of artificial intelli-
lems that they try to solve. In fact, the programmer must first gence. Springer, Cham, 2016. p. 555-572.
obtain the most representative sample possible. Then he will [17] PAN, Yunhe. Heading toward artificial intelligence 2.0. Engineering, 2016,
have to choose the function most faithful to the sample. Final- vol. 2, no 4, p. 409-413.
ly, they should be used as tools because not all problems re- [18] SEBASTIANI, Fabrizio. Machine learning in automated text categoriza-
quire a complex program in machine or deep learning. tion. ACM computing surveys (CSUR), 2002, vol. 34, no 1, p. 1-47.
Deep learning and machine learning are going to be general- [19] SNOEK, Jasper, LAROCHELLE, Hugo, et ADAMS, Ryan P. Practical bayesi-
ized "in all decision-making electronics", as in cars or planes, an optimization of machine learning algorithms. In : Advances in neural in-
they will make it possible to do without a human expert to formation processing systems. 2012. p. 2951-2959.
sort the data, since the algorithm will find its own correlations. [20] Outputs, with Relationships to Statistical Pattern Recognition,” Neurocompu-
To take the example of facial recognition, the DL algorithm ting—Algorithms, Architectures and Applications, F. Fogelman-Soulie and J.
will determine for itself whether to take into account the dif-
ference between the eyes (between the pixels) or if this infor-
mation is not decisive enough compared to others (and this is
indeed the case).

REFERENCES
[1] BEAM, Andrew L. et KOHANE, Isaac S. Big data and machine learning in
health care. Jama, 2018, vol. 319, no 13, p. 1317-1318.
[2] BIAMONTE, Jacob, WITTEK, Peter, PANCOTTI, Nicola, et al. Quantum
machine learning. Nature, 2017, vol. 549, no 7671, p. 195.
[3] BOTTOU, Léon, CURTIS, Frank E., et NOCEDAL, Jorge. Optimization meth-

IJSER © 2019
http://www.ijser.org