Machine Learning Foundations

Machine Learning is everywhere around us, even when we don't notice it. Every time a website suggests a video, filters a spam email or recognises a face, a machine learning system is working in the background.

The key idea

Machine Learning lets computers learn patterns from data instead of requiring programmers to write every rule for every situation.

Arthur Samuel's checkers program

Samuel wrote a program that learned to play checkers. It played many games, learned from experience, and its playing ability improved over time. It made better decisions without every move being programmed by hand.

Machine learning algorithms fall into a few main families. This module covers supervised and unsupervised learning in detail and introduces the others.

These techniques have advanced rapidly and are widely used in real-world applications. Practical advice on applying them matters just as much as the algorithms themselves, and we return to it throughout the course.

During training the model sees examples of both X and the correct Y. It learns the relationship between them, then uses it to make predictions on new inputs.

Where supervised learning is used

Two main types

Andrew Ng explains regression with house price prediction. Each point below is one house: X is its size in square feet and Y is its selling price. The model learns the relationship between size and price.

Linear and nonlinear regression

Linear regression models the relationship with a straight line:

• x: the input feature (house size)
• w: the weight, or slope
• b: the bias, or intercept
• f(x): the predicted output (price)

Nonlinear regression uses a curve when a straight line cannot represent the data well. A curve may fit some datasets better, but a more complex model is not automatically more accurate on new data.

Characteristics of regression

• Predicts numerical values that can take many possible values.
• Learns the relationship between features and a numerical target.
• Can use one input feature or several.
• Used for forecasting, estimation and numerical prediction.

Unlike regression, classification predicts which category an input belongs to. Examples include spam detection, recognising a handwritten digit, or predicting whether a tumour is benign or malignant.

Breast cancer example

Andrew Ng uses tumour detection as his classification example. Here is a small dataset of tumour sizes and diagnoses:

Andrew Ng extends the tumour example to three classes: Class 0 is benign, Class 1 is malignant type 1 and Class 2 is malignant type 2. Another multiclass example is sorting images into cat, dog or horse.

Regression vs classification

A classification model can use more than one input feature. In Andrew Ng's example the two features are X₁: tumour size and X₂: patient age, and the output Y is benign or malignant.

Supervised learning trains on pairs (X, Y). Unsupervised learning has only X, so the algorithm must find useful structure in the inputs by itself.

Supervised vs unsupervised

Three important unsupervised problems

K-Means

K-Means splits data into K groups. It places K cluster centres, assigns every point to its nearest centre, moves each centre to the middle of its group, and repeats until nothing changes.

• K: the number of clusters
• Centroid: the centre of a cluster
• Data point: one observation
• Distance: how far apart two observations are

Application 1: grouping news stories

Andrew Ng shows how news articles about the same event can be grouped. The wording differs, but all four headlines describe one event, so a clustering system can place them together.

Application 2: DNA microarray

A DNA microarray measures gene expression. Each row is a gene, each column is a person or sample, and the colour shows the expression level. Clustering finds samples with similar profiles, which helps researchers study disease subtypes and biological patterns.

Application 3: customer segmentation

Customer segmentation uses clustering to discover groups of customers with similar needs. Andrew Ng's example finds three kinds of interest:

A training institute could use course interests and enrolment behaviour in the same way, to see which students prefer programming, career development or AI courses.

Suppose a customer usually spends between ₹500 and ₹2,000 per transaction. One day a transaction of ₹2,00,000 appears. An anomaly detection system can flag it for review because it is far from the customer's usual spending.

Where it is used

• Banking fraud investigation
• Cybersecurity
• Manufacturing quality monitoring
• Healthcare monitoring
• Unusual website traffic detection

Suppose a dataset has 100 features. A dimensionality reduction method might describe it with just 10 new features.

Benefits

• Lower computational cost and faster model training
• Easier visualisation
• Less redundancy
• Potentially less noise

PCA: Principal Component Analysis

PCA is a popular dimensionality reduction technique. It builds new variables called principal components that capture the directions in which the data varies the most. Below, two features are squashed onto one line. The best line keeps as much of the spread as possible.

Jupyter Notebook is an interactive environment for writing, running and documenting Python code. It is widely used for machine learning, data science, data visualisation, numerical analysis and scientific experiments.

Optional labs

Andrew Ng's optional labs help you reinforce the lectures, understand working code, gain hands-on experience and prepare for graded programming exercises. You can often complete them just by running and reading the existing code cells.

Two kinds of cells

# Machine Learning
This notebook introduces Python.

print("Hello, Machine Learning!")

Hello, Machine Learning!

Running code

You can also run cells with the notebook toolbar.

An f-string lets you put variables and expressions inside a piece of text. Write an f before the opening quote and wrap the value in curly braces.

name = "Andrew"
print(f"Hello, {name}!")

Hello, Andrew!

a = 7
b = 5
print(f"{a} + {b} = {a + b}")

7 + 5 = 12

Practical Lab 1: Python and Jupyter Notebook exploration. Duration: 60 minutes, split into four sessions.

Session 1: menu discovery (10 minutes)

Explore the main menus of classic Jupyter Notebook.

Session 2: Python fundamentals (15 minutes)

Practise variables, data types, arithmetic, lists, functions and f-strings. Click a cell and press Run (or Shift + Enter) to see its output.

Exercise 1: variables and f-strings

name = "Machine Learning Student"
course = "Machine Learning"
hours = 1

print(f"Name: {name}")
print(f"Course: {course}")
print(f"Practice duration: {hours} hour")

Name: Machine Learning Student
Course: Machine Learning
Practice duration: 1 hour

Exercise 2: lists and functions

prices = [100, 200, 300, 400]

def calculate_average(values):
    return sum(values) / len(values)

average = calculate_average(prices)
print(f"Average price: {average}")

Average price: 250.0

Session 3: NumPy fundamentals (15 minutes)

Practise creating arrays, indexing, mathematical operations, mean, minimum and maximum.

Exercise 3: NumPy arrays

import numpy as np

numbers = np.array([10, 20, 30, 40, 50])

print("Array:", numbers)
print("Mean:", np.mean(numbers))
print("Maximum:", np.max(numbers))
print("Minimum:", np.min(numbers))
print("Doubled:", numbers * 2)

Array: [10 20 30 40 50]
Mean: 30.0
Maximum: 50
Minimum: 10
Doubled: [ 20  40  60  80 100]

Session 4: experimentation (20 minutes)

Run cells, change values, watch the outputs, create and correct errors, restart the kernel and run the whole notebook again.

Exercise 4: understanding errors

number = 10

try:
    result = number / 0
except ZeroDivisionError as error:
    print(f"Error detected: {error}")

Error detected: division by zero

This shows how Python reports an error and how an exception can be handled instead of stopping your program.