Introduction to AI (Artificial Intelligence),

ML (Machine Learning) and Applications

Dr.Ashok Rao
Former Head, Network Project, CEDT, IISc, Bangalore

< ashokrao.mys@gmail.com >

 Outline of the talk
• This presentation is in TWO parts:
• Part-1
• Understanding concepts of AI-ML-and related ideas of DS and IoT in
the context of applications. (What, How and When of AI-ML)
• Part-2
• Application of various aspects of AI and ML in Several Domains.
(Details of how AI-ML is used in different domains with remarkable
success)
Terminologies: AI vs. ML vs. DL vs. DS
Artificial intelligence (AI)
 It is a broad term that refers to the process of making machines to
simulate the human brain function.

Machine Learning(ML)
 A Subset of AI.
 In machine learning, statistical methods are used to empower
machines to learn from data instances without being programmed
explicitly.
 The machine learns itself how to behave when a new data instance is
provided.
Terminologies: AI vs. ML vs. DL vs. DS
Deep Learning(DL)
 It is a subset of machine learning where artificial neural networks(
algorithms inspired by the human brain functioning), learn from
large amounts of data.
Data Science(DS)
 Refers to the process of extraction of useful insights from data.
 This interdisciplinary approach merges various fields of computer
science, scientific processes and methods, and statistics in order to
extract data in automated ways.
Real Data is central to ML, DL and DS
Relation between: AI,ML,DL,DS
More Pedagogic view of AI and its Avatar’s

Notice
The
Position of
DATA
GENESIS OF AI-ML - CHECKER GAME (BOARD GAME)

10
 ARTIFICIAL INTELLIGENCE

Dartmouth Hall Commemorative Plaque., Courtesy of James Moor

11
 ARTIFICIAL INTELLIGENCE

“To proceed on the basis of the conjecture

that every aspect of learning or any other
feature of intelligence can in principle be
so precisely described that a machine can
be made to simulate it.”

Dartmouth Hall Commemorative Plaque., Courtesy of James Moor

12
• Dr. Claude E. Shannon
demonstrating to Chessmaster
Edward Lasker his (home-made)
electric chess automation, build in
1949. The machine could handle
up to six pieces, and was
designed to test various
programming methods. With one
hundred and fifty relay operations
required to complete a move, it
arrived at the reply to an
opponent's play in ten to fifteen
seconds. It had built into it a
random element, and as a result
did not necessarily always make
the same move when faced with
the same position.
 EVOLUTION OF AI

IBM Deep
blue
defeated
Garry Kasparov
on
10th. Feb. 1996
By James the photographer - https://www.flickr.com/photos/22453761@N00/592436598/, CC BY 2.0,
https://commons.wikimedia.org/w/index.php?curid=3511068

14
 EVOLUTION OF AI

IBM Watson defeated two human contestants

in the show “Jeopardy“ (a quiz show)

By Source, Fair use, https://en.wikipedia.org/w/index.php?curid=31142331

15
 EVOLUTION OF AI

• Go is an abstract
strategy board game to
surround more territory
than an opponent
• Require multiple layers
of strategic thinking

https://en.wikipedia.org/wiki/Go_(game)

16
AlphaGo
DEEPMIND

18
 AlphaGo is the first computer
AlphaGo program to defeat a Go world
champion in 2015

Training using ML computer approach : STRATEGIC

 What ML can do?
Some of the popular applications of ML that are proven to be successful:
• Self-driving cars
• Medical diagnostics
• Smart Farming
• Climate Prediction
• Genome Analysis ( Gene Therapy and Gene Lineage, Angelina Joe)
• Astronomic data analysis
• Financial transaction analysis and anomaly detection (Fin Tech)
• Cyber security threat detection
• Supply Chain Management (Block Chain Tech.)
 Medical diagnostics
• Machine learning can help
in early diagnosis and
treatment decision of
many deadly diseases like
cancer.
• Potential to reduce the
workload of radiologist
and pathologists to a great
extent.
• Potential to make
diagnostic procedures a lot
more affordable, accurate
and accessible.
Self-driving cars
• self-driving car is a vehicle that is capable of sensing its environment
and moving safely with little or no human input
Self-driving cars – The Indian Scenario
• Our traffic management system and infrastructure have to leap-frog
before implementing self-driving cars in India.

• Nevertheless, We should not support this as it will cost a minimum

of 25 Lakh Jobs and Livelihood of 1 crore people.

• Every Technology has Good and Bad aspects. We

have to understand this and choose and deploy
carefully.
• This involves Soci-Economic-Context-Culture-
Social Impact, Equity-Equality etc.
(Customised Domain in India) Judiciary system
● Adjectives - Corrupt. Biased. Broken. Failed
● More than 3 crore cases pending in different parts of
India, lot of these for more than 10 years.
● 5-8 lakh cases being lodged everyday
● Global standard - 20 judges for 10 lakh people, India has
a huge gap
 Future
● Predicting outcome of a legal case using Machine
Learning algorithms.
● Understand the likelihood of winning a given case.
● Deloitte - 1,00,000 jobs in the legal sector have a high
chance of being automated in the next twenty years
AI in Indian Judiciary system: Lawyers will object,
Clients will be happy , except LPY and their likes
 Data Analysis and Data Analytics
 Turning raw data into useful information
 Purpose is to provide answers to questions being asked at a
program site or research questions
 Even the greatest amount and best quality data mean nothing if
not properly analyzed—or if not analyzed at all
 Analysis does not mean using computer software package
 Analysis is looking at the data in light of the questions you need to
answer:
 Analytics is getting beyond “Statistical” Insights of the data.
Useful in Predicting the “Behavior” of the process (Machine, Fin).
ML-Data Analytics in Domain of Smart Grid
 Use Cases of Data Analytics in Healthcare

Medical Image Analysis

 The deep-learning
based algorithms
increase the diagnostic
accuracy by learning
from the previous
examples and then
suggest better treatment
solutions.
 Data Analytics via ML : Genetics and Genomics

 The Deep Genomics

made a remarkable
impact on predicting
the molecular effects of
genetic variation
essential to DNA
interpretation.

Source : https://www.deepgenomics.com/
 ML in Creation of drugs & Vaccines
 The drug discovery process is highly
complicated and involves many
disciplines.
 On average, it takes twelve years to get
a drug officially submitted.
 The data science and machine learning
algorithms simplify and shorten this
process, adding a perspective to each
step from the initial screening of drug
compounds to the prediction of success
rate based on the biological factors.
 Such algorithms can forecast how the
compound will act in the body using
advanced mathematical modelling and
simulations instead of the “lab
experiments”.
 ML in Predictive Medicine: Prognosis and
Diagnostic Accuracy
 The data science predictive analytics
methods learn from historical data and
make accurate predictions about the
outcomes.
 They process the patient data, make
sense of clinical notes, find the
correlations, associations of
symptoms, familiar antecedents, habits,
diseases, and then make predictions.
 Common cases include the prognosis
of disease progress or prevention to
reduce the risk and the negative
outcomes.
 The main benefit is the improvement of
the quality of life for patients and the
quality of working conditions for
doctors.
Why ML works well: Diversity of Approaches
How Does it Work?
Typical Scheme for Supervised ML
Difference Between ML and DL
Why Deep Learning Performs better?

40
 Typical answers to real life problem's

Input Output
Label: “Motorcycle”
M
images/video Suggest tags
L
Image search
…

M Speech recognition
audio L Music classification
Speaker identification
…

Web search
M Anti-spam
text L Machine translation
…

Most real life perception problems are

PATTERN RECOGNITION problems

Note: The Relevance and Context of Domain Knowledge

 How it all maps to actual problem solving?

Real World (Where all the problems come from)

Problem

Data Collection +Feature detection + Feature Selection

Selection of a Clustering Method Selection of a Classifier Model

Cluster the data Training

Label the Testing

data

No
Result
OK?
Solution
Real World (Where the solutions have to be returned)

Observe: the role of Features and not raw Data.

 Learning seems to be critical difference
between Human and Machine Perception

Supervised ANN: Back propagation, Auto encoders, Boltzman Machines, Restricted BM

Classifiers: FLD, Naïve Bayes Classifier, SVM, Linear regression
Decision Trees: c4.5, Random forest, etc.

Unsupervised ANN: Self-organizing map

Hierarchical Clustering, Partitional Clustering

Semi-supervised Graph based methods, Generative Models

Reinforcement Monte Carlo, Temporal Difference Method, Learning Automata

Deep Belief networks, Deep Boltzman Machines, Deep

Deep Learning
Convolution neural networks, Deep recurrent neural networks

There are different ways and models of learning!!

TWO extremes: Arjuna and Ekalavya
Generic ML/DL approach
 Where does it move to ML
• Machine Learning involves not Data Processing Per-se, But
• Data Processing Via Pattern Recognition or Pattern Extraction
in a cloud of data.
Pattern Extraction is done by Identifying Subtle / Obvious
Features inherent in all Data.
These are used finally to classify Data to 2 classes
(Supervised) which can be extended to n-class.
Else it can be clustered (Unsupervised) when class labels are
not available.
ML as a classification problem via Pattern Classes
 Why is this hard?

You see this:

But the camera sees this:

 Pixel-based representation

Learning
algorithm

Input
Motorbikes
Raw image “Non”-Motorbikes
pixel 2

pixel 1
 Pixel-based representation

Learning
algorithm

Input
Motorbikes
Raw image “Non”-Motorbikes
pixel 2

pixel 1
 Pixel-based representation

Learning
algorithm

Input
Motorbikes
Raw image “Non”-Motorbikes
pixel 2

pixel 1
 What we want

handlebars

Feature Learning
wheel representation algorithm
E.g., Does it have Handlebars? Wheels?
Input
Motorbikes
Raw image “Non”-Motorbikes Features
pixel 2

Wheels
pixel 1 Handlebars
Unsupervised feature learning

Find a better way to represent images

than pixels in an unsupervised manner
The goal of Unsupervised Feature Learning

Unlabeled images
Automatically

Learning
algorithm

Feature representation
Change in Programming style/Languages
Perceptron
Neural Network
Trained NN for two class problem
New basis for Computing: ANN’s

An artificial neuron is an imitation of a human neuron

 How do ANNs work?
• Now, let us have a look at the model of an artificial neuron.
 How do ANNs work?
Not all inputs are equal
............
xm x2 x1
Input
wm ..... w2 w1
weights

Processing ∑ ∑= X1w1+X2w2 + ….+Xmwm =y

Output y
Deep learning

object models

object parts
(combination
of edges)

edges

Why feature
pixels
hierarchies?
CNN - Convolutional Neural
Network
GoogleNet
RNN – Recurrent Neural
Network
Reinforcement Learning
Example of Character (numeral) Recognition
Deep Learning
 Why AI-ML Succeeds??

• Computation.
• Training (DATA)
• Model captures Phenomena.
• Most Important: DATA-DOMAIN-MODEL links, reinforcement, validity.
• So very important for using it is : DOMAIN KNOWLEDGE.

72
AI in Healthcare
Facts of Diabetes: Ground Truth
We eventually want something like this :
Where AI-ML sits in this Domain ?
Simple Explanation of the Issue
Larks AI coach for Diabetes management
Another Application in DIGITAL HEALTH
ML and AI in cardiovascular disease analysis
 Analysing Cardiac condition and
Attempt to prevent cardiac arrest
Ultimate Possibility : AI-ML in ICU can predict which patient is likely to die !!
Another Proof of Success: One of the reasons for Covid Vaccine in 2 yrs (not 12)
How to get Data: Critical Role of IoT, Embedded Systems & Sensors
The all Important item in AI-ML-DS etc
 Future
wearable
healthcare
systems.

IoT Enabled
Sensors for
Medical
Application
Surprising: Even functional items will gather data for ML-DS-DA
From this to that soon: Wearable Technologies
 Multifunctionality:
IoT Embedded Dress/ Fabric
How a typical scheme / Architecture looks like
Endorsement: Greatest teacher and his latest book
What is well established and popular (Jan 2019) !!
World class : Something from close Proximity
Many other Basic Materials that have shaped
Understanding AI in the big picture
AI and IoT for Av’s
DL applications in Autonomous Vehicles
DL for AV’s
DL for AV’s –contd.
So much of AI in AV’s
Almost all components for “safe” AV navigation
AI in Education
AI in Education cont.
IoT and AI in Agriculture
The Secret is
Sensors.

Sensors
means IoT.

Intelligence
means AI
AI has ML in
it.
AI in Defence and Weapons
AI-ML in defence
What is Next in Data Sciences: A Little Beyond
• What is “Better” than ML for Data Sciences?
• Next step in Data Sciences, Higher in Hierarchy?
Extracting Patterns VS Building Causal Models
Where ML may fail??
 Two Examples that captures
Power of AI-ML these well: First
One Example that
captures all that is
said so far.
A Predictive and
Anticipatory
Health Care:

AI-ML,Data Sciences
and Data Analytics.

Smart Hospital Bed

 Thank you
•Please understand AI-ML is not CS= C/C++ programming
•AI-ML is not = Python Programming.
•Role of Domain Knowledge is VERY VERY important.
•Build Strong Analytical , Statistical Skills and Knowledge of that.
• Do value DATA and learn from “Data”.
• Encourage students to pick up 2-3 domain knowledge while they
study any UG/PG program.
• Get Interdisciplinary, move on to Multidisciplinary outlook and
capability.
• Understand that in 5-10 yrs ML-DL may give way to something else.