Lecture 4 - Basics of ML

Prof. Ankit Gangwal

Assistant Professor, IIIT-H, India
email: [email protected]
Web: CiaoAnkit.github.io

Introduction Prof. Ankit Gangwal 93

Announcements

Introduction Prof. Ankit Gangwal 94

Table of Contents

● What is ML?
● Why ML?
● Basic Workflow of ML
● Basic Terminology in ML
● Tasks involved in Supervised Learning
○ Regression
○ Classification
● Perceptron
● Activation Functions
● Multi-layer perceptron

Introduction Prof. Ankit Gangwal 95

Table of Contents

● Learning parameters of a neural network

● Loss Functions in neural networks
○ Mean square error
○ Cross entropy
● Backpropagation, mini-batch training, SGD,
● Evaluating ML models
● Overfitting vs Underfitting
● Early Stopping
● Convolutional neural nets
● Vision CNN models for Image classification
● Miscellaneous ML topics

Introduction Prof. Ankit Gangwal 96

Basics of Machine Learning

What is Artificial Intelli ence?

Introduction Prof. Ankit Gangwal 97

Basics of Machine Learning

What is Machine Learnin ?

Introduction Prof. Ankit Gangwal 98

Basics of Machine Learning

What is Machine Learnin ?

Machine learning (ML) is a field o study in artificial intelli ence
concerned with the development and study o statistical al orithms
that can learn rom data and eneralize to unseen data, and thus
per orm tasks without explicit instructions.[1]

Introduction Prof. Ankit Gangwal 99

Basics of Machine Learning

Deep Learning??

Introduction Prof. Ankit Gangwal 100

Basics of Machine Learning

Introduction Prof. Ankit Gangwal 101

Basics of Machine Learning

AGI??

Introduction Prof. Ankit Gangwal 102

Basics of Machine Learning

AGI??

ANGI??

Introduction Prof. Ankit Gangwal 103

Why do we study ML?

● Because hard-coding is not always feasible

● We do not have an pattern algorithm for every problem
● Because when intelligently designed they can be scalable
● Adaptibility - Reinforcement Learning
● Can be applied in almost every area!
○ Healthcare - Biology, Microbiology
○ Finance
○ Robotics(Driving,Drones)
○ NLP(Chatgpt,claude)
○ Computer Vision
○ Graphs(2d and 3d)

Introduction Prof. Ankit Gangwal 104

Why do we study ML?

Introduction Prof. Ankit Gangwal 105

Basics of Machine Learning

Introduction Prof. Ankit Gangwal 106

Basic Workflow in ML

What is an ML model?

An ML model can be thought of as a function F(x).

Every function has an input and an output

Take an example of recommending movies on netflix.

In this scenario, input is the history of all the previous
movies watched by you. Output would be recommending a
movie you like. So, ML is basically trying to find the best
possible estimate for F(x).

Introduction Prof. Ankit Gangwal 107

Basic Workflow in ML

How do we find the best estimate of F(x)?

Ans: We learn!

How do we learn?
Ans: By Training

Sounds Familiar?
This is exactly how we humans learn anything

Introduction Prof. Ankit Gangwal 108

Basic Workflow in ML

Let’s take the example of humans learning how to drive..

Day 1: Terrible

Day 2: Decent This is exactly how

Day 3: Good ML models learn!!
Day 4: Expert

Introduction Prof. Ankit Gangwal 109

Basic Workflow in ML

Now let’s try to find the most important aspects of ‘learning how to drive’

Task/Goal: Learn to drive Task

Medium to learn: Driving Instructor Algorithm

Experience: longer you learn, better you get Data

This is literally how you train an ML model as well!

Introduction Prof. Ankit Gangwal 110

Basic Workflow in ML

Now, let’s try to build a model which outputs 1 if there

exists the word ‘Chubby cat’ in a given paragraph…

Now, let’s try to build a model which outputs 1 if there

exists a ‘Chubby cat’ in a given image…

Introduction Prof. Ankit Gangwal 111

Some Terminology

Now, let’s define some terminology before we explore more

interesting topics…

Introduction Prof. Ankit Gangwal 112

Taxonomy of Data

Introduction Prof. Ankit Gangwal 113

Taxonomy of ML Tasks

Introduction Prof. Ankit Gangwal 114

Regression

Predicts continuous numerical values.

Ex: Stock Market Prediction

Housing price prediction
Estimating Blood sugar
Estimated Delivery time in e-commerce

Introduction Prof. Ankit Gangwal 115

Classification

Categorizes data into classes or groups

Ex: Image Classification
Video Classification
Hate Speech Detection
Link Prediction in social networks
Recommendation system

Introduction Prof. Ankit Gangwal 116

Building a Neural network

● Now, our goal is to build a mathematical function F(x) such that

it is learnable and can process input.
● So, we basically want to mimic the human brain.
● We start with the smallest unit of the human brain.

Introduction Prof. Ankit Gangwal 117

Neuron

3 main parts:
● Dendrites
● Soma
● Axon terminal

This is similar to a function.

● Input
● Processing
● Output

Introduction Prof. Ankit Gangwal 118

Neuron

We can start with the simplest mathematical function which can take an
input, process it and produce an output.

This is obviously not enough complexity to learn complex equations.

Let’s try to list out some of the problems.

1. Multiple input features

2. Linearity

Now, let’s try fix these problems one by one.

Introduction Prof. Ankit Gangwal 119

Neuron

Before we fix the problems of Neuron, let’s try to visualize what it does.

Introduction Prof. Ankit Gangwal 120

Multiple Input features

This is an easy fix. We can simply assign weights to every single feature.

We can simplify this representation by using matrices.

Introduction Prof. Ankit Gangwal 121

Multiple input features - Visualization

Introduction Prof. Ankit Gangwal 122

Linearity

The capabilities of non-linear equations is limited.

To mitigate this issue, we introduce activation functions.

An activation function is a mathematical function applied to the output of a

neuron. It introduces non-linearity into the model, allowing the network to
learn and represent complex patterns in the data.

Examples include:
● Linear
● Sigmoid
● Tanh
● ReLU
● GeLU
● Softmax etc..

Introduction Prof. Ankit Gangwal 123

Activation Functions
Linear Activation Function

Introduction Prof. Ankit Gangwal 124

Activation Functions
Sigmoid Activation Function

Characterized by it’s S shape.

Output is restricted to between 0 and 1.

Introduction Prof. Ankit Gangwal 125

Activation Functions
Tanh Activation Function

Shifted version of sigmoid that is stretched along the y-axis

Output is between -1 and 1.

Known for it’s 0-centered distribution

Introduction Prof. Ankit Gangwal 126

Activation Functions
ReLU Activation Function

Output range is between 0 and -inf

Introduction Prof. Ankit Gangwal 127

Neural Network
Now, let’s try to piece all of them together to create the most basic form of a neural
network that can model non-linear functions

Introduction Prof. Ankit Gangwal 128

Deep Neural Network

Introduction Prof. Ankit Gangwal 129

Deep Neural Network (Multilayer Perceptron)

Why do we need deeper neural networks?

Introduction Prof. Ankit Gangwal 130

Universal approximation theorem

In simple words, the universal approximation theorem says that neural

networks(with at least one hidden layer) can approximate any function.

Introduction Prof. Ankit Gangwal 131

Deep Neural Network

Now, let’s try to build a simple feedforward neural network with

multiple hidden layers.

Introduction Prof. Ankit Gangwal 132

Forward Propagation

Introduction Prof. Ankit Gangwal 133

Forward Propagation

Introduction Prof. Ankit Gangwal 134

Forward Propagation

Introduction Prof. Ankit Gangwal 135

Forward Propagation

Introduction Prof. Ankit Gangwal 136

Forward Propagation

Introduction Prof. Ankit Gangwal 137

Forward Propagation

Introduction Prof. Ankit Gangwal 138

Forward Propagation

Introduction Prof. Ankit Gangwal 139

Forward Propagation

Introduction Prof. Ankit Gangwal 140

Forward Propagation

Introduction Prof. Ankit Gangwal 141

Designing Neural Networks
How do you design an MLP for Regression task?

How do you design an MLP for Classification task?

Introduction Prof. Ankit Gangwal 142

Training a Neural Network
Now, we built a function that is capable of acting as an approximator for any
continuous function when you assign appropriate weights to it

How do we find these weights?

Revisiting how humans learn a task…

Whenever we make a mistake, we try to learn from it and adjust ourselves

accordingly.

For MLPs also we need to define a metric that shows how far off we are from
the actual output

Introduction Prof. Ankit Gangwal 143

Loss Functions

Introduction Prof. Ankit Gangwal 144

Loss Functions

Introduction Prof. Ankit Gangwal 145

Loss Functions
Regression

● Mean Square Error

● Mean Absolute Error

● Relative Mean Square Error

Introduction Prof. Ankit Gangwal 146

Loss Functions
Classification

● Binary Cross entropy loss

● Cross entropy loss

Introduction Prof. Ankit Gangwal 147

Loss Functions

● Now, we have defined the metric for error.

Introduction Prof. Ankit Gangwal 148

Loss Functions

● Now, we have defined the metric for error.

● Next, we have to come up with an algorithm such that weights can be adjusted
in order to decrease this error.

Introduction Prof. Ankit Gangwal 149

Loss Functions

● Now, we have defined the metric for error.

● Next, we have to come up with an algorithm such that weights can be adjusted
in order to decrease this error.
● We can now rewrite this as an optimization problem.

Introduction Prof. Ankit Gangwal 150

Thank you!

Introduction Prof. Ankit Gangwal 151