Chapter 10

Machine Learning Applications

Dr.K.Rajkumar
Dean – International Relations
Associate Professor of CS
Machine learning should be considered when:

Complexity: The rules are too complex or there are too many rules
Scalability: The data volumes are too large for traditional data
mining methods

• Machine learning works well if there is a lot of data and if there

is a ground truth.
• Recent technological developments, such as the advent of the
Internet of Things (IoT), mobile devices and cloud computing,
are likely to produce an ever increasing amount of data.
• Machine learning provides the functionality to effectively
analyze the volumes of data produced and it is to be expected
that the importance of machine learning will increase in the
future.

Dr.K.Rajkumar, Dept.of CS, BHC 2

 Anomaly detection
• Anomaly detection refers to a wide range of unusual activities,
such as fraudulent bank transaction, network security breaches
or abnormal sensor data suggesting imminent machine failure.
• The increasingly digitized world and the advent of Internet of
Things produces an ever-increasing amount of transactional
and sensor data.
• The main goal of digitization is to increase productivity and
efficiency and ultimately reduce costs.
• There are different methods to identify abnormal data points.
• In some simple cases, data visualization can be used, for
instance, to detect outliers that are located outside the normal
data distribution.

Dr.K.Rajkumar, Dept.of CS, BHC 3

 • The easiest way to detect anomalies is to verify that certain
values, such as sensor data measurements, are out of bounds.
• If some threshold is exceeded, an alarm is sent.
• The problem with this approach is that there is the risk of
having a high false positive rate or that some anomalies go
undetected.
• Usually, we cannot consider single sensor values on their own
when determining the health of an equipment, we need to take
into account combinations of sensor values, in other words,
patterns.

Dr.K.Rajkumar, Dept.of CS, BHC 4

 Security
• One problem with anomaly detection is that an anomaly might be
unseen and there is no historic data available that reflects the
anomaly.
• For instance, there are constantly new security threads detected
that have been unknown and, hence, a machine learning algorithm
cannot be trained on existing data.
• A zero-day vulnerability is such an unknown or unaddressed
security vulnerability that can be exploited by hackers.
• If it detects some deviant traffic, it issues an alarm and a security
specialist can analyze the suspicious traffic.
• When we have a collection of data points from network traffic
they typically have a certain distribution, such as a Gaussian
distribution.

Dr.K.Rajkumar, Dept.of CS, BHC 5

 Steps used for Anomaly Detection
• First calculate the probability distribution p(x) from the data
points.
• For every new data point x, the probability that it belongs to
the probability distribution is calculated and compared against
a threshold.
• If p(x) is smaller than the threshold, x is considered an
anomaly.
• Normal data points tend to have a large p(x), whereas
anomalous data points tend to have a small one.

Dr.K.Rajkumar, Dept.of CS, BHC 6

 Predictive Maintenance
• Predictive maintenance is used to monitor the health state of an
equipment.
• Modern machines are equipped with sensors that collect data,
such as the temperature, humidity, vibration or other values,
depending on the type of equipment.
• The early detection and avoidance of technical problems is also
called predictive and preventive maintenance.
• Smart logistics and preventive maintenance are two main
application fields and drivers of Internet of Things.
• One solution for predictive maintenance is to use the Anomaly
Detection.
• Another approach uses an Autoencoder.

Dr.K.Rajkumar, Dept.of CS, BHC 7

 • The autoencoder creates a lower dimensional representation of
the sensor values with their correlations and interactions and
reconstructs them back to the original values.

• The autoencoder is trained on normal sensor data that

represents the normal operating state.

• If the monitored equipment degrades, the interactions between

the variables change, which results in an increase of the error
of the reconstructed input values.

• We use the probability distribution of the reconstruction

error in order to determine if a data point is normal or if we
have an anomaly.

Dr.K.Rajkumar, Dept.of CS, BHC 8

 Biomedical Applications
Medical applications
• Machine learning has found a large application field in pharma
and medicine.
• It is used to support physicians in disease identification and
diagnosis, for medicine discovery to find new drugs or for
personalized treatments.
• Machine learning methods and, in particular, convolutional
neural networks (CNN) have been successfully applied to
analyze medical images such as CAT scans, X-rays and MRI
scans.
• Convolutional neural networks have been used for early
detection of illnesses such as Alzheimer’s or fibrosis.
• They are also used to classify images or tissue samples into
malign or benign tumors.

Dr.K.Rajkumar, Dept.of CS, BHC 9

 • Data from a health tracker, such as blood pressure or blood
sugar levels, can also be transmitted.
• Machine learning can significantly improve healthcare by
allowing patients who need to regularly see a doctor or who
need stationary treatment to be monitored remotely, allowing
them to live a more independent life.
• Machine learning is a major enabler of personalized treatments
since it has the capability to analyze large amounts of data, but
also to correlate different types of data such as leukocyte
levels and genetic information.
• Using predictive analysis for disease assessment and
management helps physicians to select the right treatment and
adapt it quickly if some values have changed.

Dr.K.Rajkumar, Dept.of CS, BHC 10

Natural Language Processing
• Natural language is the language humans speak and write and the
attempt to extract information from texts or speeches using algorithms
is called natural language processing or NLP.
• Natural language processing encompasses understanding human
language, e.g., understanding voice commands on a Smart Phone, and
generating natural language, e.g., a text to speech application that
reads Web pages to people with visual impairments.

Dr. K. Rajkumar, Dept of Computer Science 11

Natural Language Processing
History of NLP:
• Natural language processing dates back to the beginnings of machine
learning in the 1950s, when Alan Turing wrote a seminal paper called
“Computing machinery and intelligence”.
• In this paper, Turing proposes an imitation game where an interrogator
needs to figure out the gender of a man and a woman invisible to him, by
asking them questions.
• If one of the invisible persons is replaced by a computer and the interrogator
cannot determine anymore if he is talking to the human or the computer, the
computer can generate human cognitive capabilities.
• This is now called the Turing test.

Dr. K. Rajkumar, Dept of Computer Science 12

Natural Language Processing
• Machines are capable of understanding spoken commands or
determine the subject of a text, but, to humans, a sentence or text has
meaning, to a machine it is an array of characters.
• A machine does not understand the context what makes word sense
disambiguation a tricky task.
• The word “meeting” can be a noun, as in “we are in a meeting”, or a
verb, as in “we are meeting an my office”.

Dr. K. Rajkumar, Dept of Computer Science 13

Text Mining
• Text mining or text analysis, aims to extract information from texts
automatically.
• The applications of text mining are innumerable.
• Text mining is used for text categorization, text clustering, document
summarization or opinion mining, but many other use cases exist.
• Spam filtering which is an application of text classification where the
emails are classified into legitimate and into spam mails.

Dr. K. Rajkumar, Dept of Computer Science 14

Text Mining
Sentiment Analysis
• Sentiment analysis, also called opinion mining, is the field of study
that analyzes peoples opinions, sentiments, evaluations, appraisals,
attitudes, and emotions towards entities such as products, services,
organizations, individuals, issues, events, topics, and their attributes.
• Sentiment analysis is extensively used by companies to determine the
popularity of products or services in order to improve them.
• It is also used in politics to make predictions on the outcome of
elections or to determine the popularity of TV programs or series.

Dr. K. Rajkumar, Dept of Computer Science 15

Text Mining
Sentiment Analysis
• If a text contains many positive sentiment words or more positive than
negative sentiment words, it is considered a positive review.
• However, the bag-of-words approach considers only individual words
and cannot handle sentiment polarity shifters.
• A sentiment polarity shifter changes the meaning of an expression to
the opposite.
• There is more sophisticated approach that uses n-grams to capture the
relationship between words.

Dr. K. Rajkumar, Dept of Computer Science 16

Text Mining
Sentiment Analysis
• A tokenizer is used to split a text into sequences of words.
• If we have sequences of two words, they are called bigrams, and we
analyze how often a word X is followed by word Y.
• One approach is now to count bigrams and do frequency analysis on
the bigrams. If we have many bigrams such as “not great” or “very
bad” the text is considered a negative review.
• When using a shallow learner, such as the multilayer perceptron, we
have to extract the features manually.

Dr. K. Rajkumar, Dept of Computer Science 17

Text Mining
• A feature might be a part-of-speech tag on a word.
• In part-of-speech tagging, a word is assigned in a text to a particular
part of speech based on its relationship with adjacent and related
words in a text.
• A word is identified as noun, verb, adjective, etc., for word-category
disambiguation.
• For instance, the word “fly” can mean the insect, when it is a noun, or
“to fly”, when it is a verb.

Dr. K. Rajkumar, Dept of Computer Science 18

Text Mining
• After tokenization and part-of-speech tagging, the word stem, if found,
and the words are grouped according to their word stems, a process
called lemmatization.
• The words “had”, “has” and “having” are grouped since they have the
same word stem, “have”.
• Lemmas extend the feature set by adding word groups based on stems.
• Using deep learners, these steps, such as lemmatization and part-of-
speech tagging, are done automatically, which is a big advantage over
shallow approaches.
• However, deep learners require a large amount of training data that is
not always available.
Dr. K. Rajkumar, Dept of Computer Science 19
Other Applications
• Machine learning has many more applications, but some notable area
of utilization is the military.
• Drone surveillance produces a tsunami of data that needs to be
analyzed quickly before, for instance, a potential enemy has
disappeared.
• The Department of Defence has specifically established an
Algorithmic Warfare Cross-Functional Team (AWCFT) in order to
integrate artificial intelligence and machine learning more effectively
across operations so as to maintain advantages over increasingly
capable adversaries and competitors

Dr. K. Rajkumar, Dept of Computer Science 20

Other Applications
• Using AI, Project Maven will initially provide vision algorithms for
object detection, classification, and alerts, according to the memo
establishing the team.
• Google has supported the Pentagon with a special API to its Tensor-
Flow machine learning framework for analyzing images from drones.
• Google is providing AI to the Pentagon Project Maven, which aims
to speed up analysis of drone footage by automatically classifying
images of objects and people in the MQ-9 Reaper drone.
• Its goal is to make the drone attack military targets automatically.
• Ultimately, the goal is to develop autonomous systems, such as drones
and other airborne tactical areal systems, that can attack independently.
Dr. K. Rajkumar, Dept of Computer Science 21
Other Applications
• Machine learning can be used to analyze and respond to cyber attacks
as described before.
• Machine learning has been used in cyber security for detecting
anomalies in network traffic and malware.
• However, machine learning can also be used to create malware.
• Developing malware, Trojans, ransomeware or malicious scripts, is
largely a manual process for a cyber criminal.
• Machine learning can automate this process or speed it up
considerably.
• Algorithms based on generative adversarial network (GAN) have
already been proposed
Dr. K. Rajkumar, Dept of Computer Science 22
Other Applications
• More sophisticated attacks use botnets and swarm intelligence.
• A swarm is a decentralized, self-organizing system.
• Bots that are interconnected through the Internet form a botnet.
• The community of computers in a botnet transmit spam or viruses
across the web without their owners permission.
• Self-learning swarmbots form a hivenet.
• Hivenets share information to create a customized attack.
• They can, for instance, attack any access point in a company network.
• Bots can run on any device connected to the internet, smart phones,
baby phones, TV boxes or soundbars.
Dr. K. Rajkumar, Dept of Computer Science 23
Other Applications
• Internet of Things (IoT) is the connectivity of the Internet into the
physical world.
• Internet of Things produces a lot of data and is only useful when the
data can be analyzed effectively.
• The data produced by Internet of Things enabled devices can be of any
form; numeric data, streaming data, text data or binary data.
• The data is transmitted to the cloud, where it is being analyzed and
visualized for end users to be assessed.
• Internet of Things can connect all equipment in a whole production
plant to monitor its health state and determine if a device or part needs
maintenance or replacement.
Dr. K. Rajkumar, Dept of Computer Science 24
Other Applications
• The data is pre-processed and analyzed on systems close to the
physical devices and only a consolidated data set is sent to the cloud.
This is called edge computing.
• Data analysis software is deployed on the edge computer. Edge
computers are machine learning enabled and usually support deep
learning too.
• The edge computer monitors the connected devices for potential
problems using predictive and preventive maintenance.
• If it determines that an equipment needs maintenance, it sends an alert
to the cloud where a service technician can be assigned to correct the
problem.
Dr. K. Rajkumar, Dept of Computer Science 25
Other Applications
• The data mining system in the cloud analyses the consolidated
information transmitted from the different edge computers. This is
called distributed analytics.
• Distributed analytics spreads the workload aver multiple nodes, the
edge computers.
• This is called fog computing or fog networking, where a network of
connected edge computers carry a substantial amount of computing
and storage.
• This results in faster insights into the health state of the connected
devices.

Dr. K. Rajkumar, Dept of Computer Science 26

Other Applications
• Ideally, the edge computers can communicate with each other and do
collaborative machine learning
• Machine learning enabled Internet of Things solutions are not limited
to predictive and preventive maintenance.
• They can be used to optimize the production in a factory, save
resources such as energy or raw materials and reduce waste.

Dr. K. Rajkumar, Dept of Computer Science 27