Unit 2: Advanced Concepts of Modeling in AI - AI, ML, DL, Machine Learning Models & Neural Networks

Unit 2: Advanced Concepts of Modeling in AI

Building an Artificial Intelligence project requires working with models or algorithms. These models may be designed from scratch or adapted from pre-existing models. Before understanding different modeling techniques, it is important to understand the relationship between Artificial Intelligence (AI), Machine Learning (ML), and Deep Learning (DL).

---

2.1 Revisiting AI, ML and DL

The relationship between AI, ML and DL can be understood using an umbrella approach.

1. Artificial Intelligence (AI)

Artificial Intelligence (AI) is the umbrella terminology referring to any technique that enables computers to mimic human intelligence.

An artificially intelligent machine works on data and algorithms fed to it to produce a desired output.

2. Machine Learning (ML)

Machine Learning (ML) is a subset of Artificial Intelligence that enables machines to improve at tasks through experience.

ML models learn from new data and use that experience to improve their future performance, taking into account mistakes and exceptions.

Examples:

• Anomaly detection in heart-rate monitors
• Object classification in photographs

3. Deep Learning (DL)

Deep Learning (DL) is a subset of Machine Learning and represents a more advanced approach to learning from large amounts of data.

It enables software to train itself using vast amounts of data and multiple machine learning algorithms working together, often using Artificial Neural Networks (ANN).

Examples:

• Identifying a bird from pixels
• Recognizing handwritten digits

AI vs ML vs DL

Concept	Meaning	Relationship	Example
AI	Enables computers to mimic human intelligence.	Broadest concept	Intelligent decision-making system
ML	Enables machines to improve through experience and data.	Subset of AI	Spam detection
DL	Uses large amounts of data and multiple algorithms, often through neural networks.	Subset of ML	Handwritten digit recognition

---

Common Terminologies Used with Data

AI models become intelligent based on the data with which they are trained. The following terms are commonly used while working with data.

Term	Meaning	Example
Data	Information in any form, such as a table containing information about different items.	Student marks, attendance, names
Features	Columns or characteristics that describe data points.	Colour, weight, height
Labels	Tags or meanings attached to data according to the problem being solved.	Fruit type such as Apple or Mango
Labeled Data	Data that has a tag or label attached to it.	Email marked as Spam
Unlabeled Data	Raw data without tags or labels.	Images without identified categories
Training Data Set	A collection of examples given to a model so that it can learn patterns.	Solved examples given to a student before an examination
Testing Data Set	Data used to test the accuracy and performance of a trained model.	New questions given in a class test

Competency-Based Questions

---

2.2 Modelling

AI Modelling refers to the process of developing algorithms or models that can be trained to produce intelligent outputs.

Generally, there are two primary approaches for building AI models:

1. Rule-Based Approach

In a Rule-Based Approach, relationships or patterns in the data are defined by the developer.

The machine follows specific instructions or rules provided by the programmer.

How It Works

The machine is provided with data + rules. It then reacts according to those predefined rules to produce the desired output.

Example

A website FAQ chatbot may contain the rule:

Major Drawback

Rule-based learning is static. Once the rules are defined, the model does not automatically learn from changes in the data or user feedback.

If the machine receives data that differs from the predefined rules, it may fail to produce the correct output.

2. Learning-Based Approach

A Learning-Based Approach enables a computer to learn how to perform a task by looking at examples or receiving feedback, similar to how humans learn from experience.

How It Works

Instead of explicitly programming every rule, the machine is provided with data and desired outputs. It identifies patterns and develops its own rules or algorithm to connect inputs with outputs.

Adaptability

Learning-based models are adaptive. When data changes, the model can modify itself to handle new patterns and exceptions.

Examples

• Unlabeled Data: A model receives 1,000 random images of stray dogs and identifies groups based on features such as colour, size and fur style.
• Labeled Data: A spam filter learns patterns that distinguish spam emails from legitimate emails.

Rule-Based vs Learning-Based Approach

Feature	Rule-Based Approach	Learning-Based Approach
Logic Defined By	Developer / Human	Machine / Self-learned
Input	Data + Rules	Data + Output
Flexibility	Static	Adaptive
Ability to Handle New Patterns	Limited	Can adapt to new data
Example	FAQ Chatbot	Spam Filter / Image Clustering

Competency-Based Questions

---

Categories of Machine Learning-Based Models

Learning-based approaches include Machine Learning and Deep Learning. Machine Learning models are generally divided into three primary categories:

1. Supervised Learning

Supervised Learning works with a labeled dataset. The machine learns from examples where the correct output is already known.

It is similar to a teacher providing solved examples to a student and then asking the student to solve new questions.

Goal: Determine relationships through training.

Examples:

• Identifying a coin based on its weight
• Email spam filtering
• Identifying a person in a tagged photograph

2. Unsupervised Learning

Unsupervised Learning works with an unlabeled dataset. The machine discovers patterns, relationships and trends without predefined labels.

Goal: Discover hidden patterns within data.

Examples:

• Grouping supermarket customers according to purchase behaviour
• Content recommendations
• Identifying suspicious transaction patterns

3. Reinforcement Learning

Reinforcement Learning enables a computer to make a series of decisions to maximize a reward for a task.

The machine learns through trial and error. Correct actions receive positive feedback, while incorrect actions receive negative feedback.

Examples:

• Self-driving cars learning to park
• Humanoid robots learning to walk
• AI agents learning to play games

Comparison of Machine Learning Models

Feature	Supervised Learning	Unsupervised Learning	Reinforcement Learning
Data Type	Labeled Data	Unlabeled Data	No pre-existing data required
Primary Goal	Determine relationships	Discover new patterns	Maximize rewards
Learning Path	With guidance	Without guidance	Trial and error
Nature	Guided learning	Pattern discovery	Highly adaptive

Case Study: The "New Species" Researcher

A researcher has a dataset of 1,000 images of various animals and birds but has no labels for them. The researcher wants the machine to find patterns and group them.

Q: Which learning model should be used?

Answer: Unsupervised Learning.

Reason: The dataset is unlabeled. The machine identifies features such as wings, fur or size and groups the data according to similarities.

Competency-Based Questions

---

Sub-Categories of Supervised Learning

Supervised Learning has two primary sub-categories:

1. Classification Model

A Classification Model is used when data is divided into distinct categories or classes.

The model learns from historical labeled data and assigns a new input to one of the predefined groups.

Goal: Predict a label or category.

Examples:

• Student grading: Grade A, Grade B or Grade C
• Email: Spam or Inbox
• Loan eligibility: Yes or No
• Weather: Hot or Cold

2. Regression Model

A Regression Model predicts a specific numerical value. Instead of placing data into categories, it estimates a numerical output.

Goal: Predict a quantity or continuous numerical value.

Examples:

• House price prediction
• Salary estimation
• Used car price prediction
• Exact temperature prediction

Classification vs Regression

Feature	Classification	Regression
Output Type	Discrete / Categorical	Continuous / Numerical
Goal	Sort into groups	Predict a numerical value
Example	Spam / Not Spam	₹4,50,000

School Example: Sports Day Trial

Task	Model	Reason
Decide whether a student is Selected or Not Selected	Classification	Output consists of categories.
Predict the exact 100m sprint time, e.g. 12.4 seconds	Regression	Output is a numerical value.

Case Study: School Canteen AI

A student is building two AI models for the school canteen.

• Model A: Predicts whether the canteen will run out of samosas today — Yes/No.
• Model B: Predicts exactly how many plates of pasta will be sold — e.g. 54 plates.

Answer:

• Model A → Classification
• Model B → Regression

If the model predicts whether the weather will be Rainy, Sunny or Cloudy, it is also a Classification task because the outputs are categories.

Competency-Based Questions

---

Sub-Categories of Unsupervised Learning

Unsupervised Learning works on unlabeled datasets. Two important sub-categories are:

1. Clustering Model

Clustering is the process of dividing data points into different groups or clusters based on the similarity of their characteristics.

The model identifies patterns in unknown data and groups similar objects together.

Classification vs Clustering

Classification	Clustering
Supervised Learning	Unsupervised Learning
Uses labeled data	Uses unlabeled data
Assigns data to predefined classes	Discovers groups based on similarity
Classes are already known	Groups are discovered by the model

Examples:

• Grouping animal and bird images
• Music recommendations based on similarity
• Customer segmentation

2. Association Model

Association is an unsupervised learning method used to find interesting relationships or dependencies between variables in a large database.

It analyzes historical patterns to determine the probability of one event or item occurring based on another.

Example: Market Basket Analysis

A supermarket analyzes purchase patterns and discovers that customers who buy bread often also buy butter.

Therefore, when a customer buys bread, the system may recommend butter.

Clustering vs Association

Feature	Clustering	Association
Primary Goal	Divide data into groups based on similarity.	Find relationships between variables.
Action	Groups items that are similar.	Finds items that frequently occur together.
Example	Grouping users according to interests.	Identifying products commonly purchased together.

Case Study: Music Discovery

A music application creates a Daily Mix by grouping similar songs based on musical features such as tempo and intensity. It also suggests merchandise that other fans commonly purchase together.

Q1: Which model creates the Daily Mix?

Answer: Clustering Model

Q2: Which model suggests merchandise commonly purchased together?

Answer: Association Model

Competency-Based Questions

---

2.3 Neural Networks

Artificial Neural Networks (ANNs) are advanced machine learning systems designed to process large and complex datasets efficiently.

What is a Neural Network?

A Neural Network is a system of organized machine learning algorithms loosely modeled on how neurons in the human brain behave.

Its primary purpose is to solve problems involving large and complex datasets, such as image recognition and handwriting identification.

Key Advantage

Neural networks can automatically extract useful features from data without requiring a human programmer to explicitly define every feature.

Architecture of a Neural Network

A neural network is divided into multiple layers. Each layer contains several blocks called nodes. Each node acts as a small processing unit.

Layer	Function
Input Layer	Acquires data and feeds it into the system. No processing occurs here.
Hidden Layer(s)	Performs processing and computation. There may be multiple hidden layers.
Output Layer	Provides the final result to the user.

Learning Process of a Neural Network

1. The network receives input data.
2. It produces an output.
3. The output is compared with the desired output.
4. The error is identified.
5. The network adjusts its internal weights.
6. The process is repeated until the network learns to produce an appropriate output.

Applications of Neural Networks

• Facial recognition
• Customer support chatbots
• Price prediction
• Handwriting recognition
• Image recognition

---

How Does AI Make a Decision?

A simplified model called a Perceptron can be used to understand how an AI system arrives at a decision.

Decision-making involves four important components:

1. Inputs
2. Weights
3. Bias
4. Calculation and Threshold

1. Inputs

Inputs are the different pieces of information considered by the AI.

For example, while deciding whether to go to a park, the inputs may include:

• Is it sunny?
• Do I have an umbrella?

2. Weights

Each input is assigned a numerical weight representing its importance.

Different inputs can have different weights depending on their importance during decision-making.

3. Bias

Bias (B) is an additional value added to the calculation. It can represent a preference or cautiousness in the decision.

4. Calculation

Inputs are converted into numerical values, such as:

• 1 = Yes
• 0 = No

The machine then performs a mathematical calculation:

5. Threshold Comparison

The calculated value is compared with a threshold.

Condition	Decision
Output is higher than the threshold	Yes
Output is lower than the threshold	No

---

Quick Revision: Unit 2 at a Glance

Topic	Key Point
AI	Broad field of making computers mimic human intelligence.
ML	Subset of AI that learns from experience and data.
DL	Subset of ML using large amounts of data and often neural networks.
Feature	Characteristic or column describing data.
Label	Tag or known output associated with data.
Rule-Based	Uses predefined rules created by humans.
Learning-Based	Learns patterns from data and can adapt.
Supervised Learning	Uses labeled data.
Unsupervised Learning	Uses unlabeled data.
Reinforcement Learning	Learns through rewards, penalties and trial-and-error.
Classification	Predicts categories.
Regression	Predicts numerical values.
Clustering	Groups similar unlabeled data.
Association	Finds relationships between items or variables.
Neural Network	Layered system of nodes inspired by biological neurons.
Perceptron	Simplified model used to understand AI decision-making.
Weight	Represents the importance of an input.
Bias	Additional value influencing the calculation.