Introduction to Machine Learning

The foundation of everything else in this course — understand what Machine Learning really is, how it differs from normal programming, and the vocabulary every later lecture builds on.

Machine Learning (ML) is the engine underneath every Generative AI system you will study in this course — ChatGPT, image generators, AI agents. Before we can understand generative AI, we must be rock-solid on what "learning from data" actually means. This lecture builds that base.

1.1 ML vs Traditional Programming

The single most important idea in this entire course is the shift from writing rules to learning rules from data.

The two paradigms side by side

Why rules alone fail

Consider spam email detection. Could you write if-else rules for it? You would need a rule for every scam phrase, every suspicious link pattern, every fake sender — thousands of rules, and spammers change tactics daily. A "simple" program cannot keep up. ML instead learns patterns in suspicious words, links and sender behaviour from labelled examples, and automatically adapts as you feed it new data.

Three kinds of algorithms

The lecture also names three algorithm categories you should recognise:

• Deterministic — same input always gives exactly the same output (e.g. sorting a list).
• Non-deterministic — output can vary between runs even for the same input.
• Probabilistic — output is based on probabilities; ML models are largely probabilistic, predicting the most likely answer.

1.2 AI vs ML vs Deep Learning

These three terms are not separate technologies — they are nested inside one another, like Russian nesting dolls.

The key difference: feature engineering

This is the distinction examiners love to test:

• Classic ML needs you to hand-feed features. To predict house prices you must explicitly give it "Area", "Bedrooms", "Bathrooms".
• Deep Learning extracts features on its own. Feed it raw pixels of a house photo and it automatically discovers edges → shapes → textures → objects, with no human telling it what to look for.

1.3 Types of Learning

ML is divided into four learning paradigms by what kind of data the model is trained on.

1. Supervised Learning — "learning with a teacher"

Like a student given a textbook with the answers in the back. Most common approach in practice. Both Classification and Regression are types of supervised learning.

• Example: Email spam detection — thousands of emails each labelled spam / not-spam.
• Real case: Credit-card fraud detection — labelled transactions (normal vs fraudulent) train a model to flag suspicious patterns in real time.

2. Unsupervised Learning — "finding hidden patterns"

Like exploring a new city with no map or guide. The main task here is Clustering.

• Example: Customer segmentation — group shoppers into "budget", "premium", "occasional" buyers without pre-defined categories.
• Real case: Netflix clusters users with similar viewing tastes to power recommendations.

3. Semi-Supervised Learning — "best of both worlds"

Labelling is expensive — imagine 10,000 chest X-rays where a doctor charges $100 to label each one. Solution: the computer first clusters similar X-rays, the doctor labels just a few per cluster, and those labels are propagated to the rest. This dramatically cuts labelling cost while keeping good accuracy.

• Example: Google Photos — you name a face once, and it finds that person in thousands of other photos automatically.

4. Reinforcement Learning — "learning by trial and error"

No dataset is given up front. A robot dog put in a room: walks forward → +10 reward; hits a wall → −10 penalty. After enough trials it learns to avoid walls. It generates its own data through rapid trial and error.

• Real case: DeepMind's AlphaGo learned Go by playing millions of games against itself, guided only by win/lose rewards, and beat the world champion.

1.4 Types of ML Tasks

While "types of learning" is about the data, "types of tasks" is about what kind of answer the model produces.

Classification — predicting categories

Output is one of a fixed, finite set of discrete classes/labels.

• Spam or Not-Spam · Disease or No-Disease · Cat / Dog / Bird
• Real case: Medical diagnosis systems predicting a COVID test result from symptoms.

Regression — predicting continuous values

Output is a number on a continuous scale — infinitely many possible values within a range.

• House price · Tomorrow's temperature · Car mileage
• Real case: Uber surge pricing — predicting a demand-based fare multiplier in real time.

Clustering — grouping similar data

Automatically discovering natural groupings in data with no predefined categories. This is an unsupervised task.

• Grouping customers by behaviour · Organising documents by topic.

A first taste of code

You will not be asked to build a full model in Lecture 1, but recognising the shape of an ML program helps. This is the universal scikit-learn pattern used in almost every later lecture:

# 1. Import a model
from sklearn.linear_model import LinearRegression

# 2. Prepare data:  X = inputs (features),  y = correct answers (labels)
X = [[800], [1000], [1200], [1500], [1800]]   # house sizes (sq ft)
y = [40, 60, 65, 75, 110]                     # prices (lakhs)

# 3. Create the model and LET IT LEARN the rule from data
model = LinearRegression()
model.fit(X, y)              # <-- this is "learning"

# 4. Use the learned rule to predict a NEW, unseen case
print(model.predict([[1400]]))

Notice: we never wrote a formula linking size to price. We gave examples; .fit() discovered the rule; .predict() applied it. That is machine learning in four lines.

Pick your answer, then press Check Answer. For coding questions press Show Solution. These mirror the MCQ + coding style of the real exam.

from sklearn.linear_model import LinearRegression

X = [[1], [2], [3], [4]]   # inputs (features)
y = [2, 4, 6, 8]           # labels  (this is SUPERVISED data)

model = LinearRegression()
model.fit(X, y)            # learns the rule  y = 2x

print("Prediction for x=10:", model.predict([[10]]))