Introduction to Machine Learning and Deep Learning
What is Machine Learning?
Tom Mitchell offered a more precise definition in 1998 that helps us break down exactly what happens in machine learning:
This definition gives us a framework with three essential components:
Task (T): What the model is trying to do
Experience (E): The data the model learns from
Performance measure (P): How we evaluate how well the model is doing
Popular Performance Measures
The performance measure we choose dramatically affects what kind of system we end up with - even when using the same data and algorithms! Depending on the task, we can use different performance measures:
For Regression: These measures evaluate models that predict continuous values (like prices, temperatures, etc.)
For Classification: These measures evaluate models that predict categorical outcomes (like spam/not spam, positive/negative)
Training data (also called training set) is the foundational dataset used to teach machine learning models to recognise patterns.
A training set is typically structured as a table or matrix:
Each row is a sample (also called a data point or an instance)
Each column is a feature (also called an attribute)
For binary classification problems, the label represents the target outcome: Label = 0 (no/negative) or 1 (yes/positive).
Quiz
What does each row in a training dataset typically represent? A data point or instance
What is a feature in a training dataset? A column representing an attribute like age or BMI
Training data is used to teach the model to recognize patterns and relationships.
The column “Label” represents whether the patient has diabetes (1 = yes, 0 = no)
Which of the following is an example of a feature from the example dataset? DiabetesPedigree
Which of the following is an example of a sample from the example dataset? P004
Types of Machine Learning
Supervised learning is often considered the most powerful approach in machine learning, particularly for well-defined problems where we have labelled examples. The term "supervised" refers to the learning process being guided or "supervised" by known correct answers.
Classification: When we predict categories or classes e.g. diagnoses where Label: Diabetes Present (1) or Absent (0)
Regression: When we predict continuous numerical values e.g. House Price Prediction where Label = Sale price ($350,000, $175,000, etc.)
Unsupervised learning finds structure in data without being told what to look for. Two most common tasks in unsupervised learning are clustering and anomaly detection. Have a look at the case studies below to understand them.
Clustering Case Study - Customer Grouping
Older customers who spend a lot
Middle-aged/older customers who spend a medium amount
Younger customers who spend very little
Anomaly Detection Case Study - Banking Security
The system learns what normal transaction patterns look like for each customer
Transactions that deviate significantly from these patterns trigger alerts
This helps identify potentially fraudulent activity without needing examples of fraud.
Reinforcement Learning: Have you ever learned a new skill through practice and feedback? Maybe you learned to ride a bike by trying, falling, adjusting, and trying again. Reinforcement learning works similarly - it's learning through experience and feedback.
Observe: The agent perceives the current state of the environment
Select action: Based on a policy (strategy), the agent chooses an action
Take action: The agent performs the chosen action
Get reward or penalty: The environment provides feedback
Update policy (learning step): The agent adjusts its strategy based on feedback
Iterate: This cycle continues until an optimal policy is found.
Review and Reflect
RMSE (Root Mean Squared Error) and R-squared are standard metrics used to evaluate how well a regression model fits continuous numerical data.
These three core types describe how AI models learn: with labels, without labels, or through feedback from interaction.
Supervised learning relies on labeled input-output pairs to train models to make predictions or decisions.
K-Means is an unsupervised clustering algorithm, not used in supervised learning tasks.
Unsupervised learning is used to detect structure in data without labels, such as grouping or reducing features.
Reinforcement learning uses trial and error, where agents learn through feedback based on actions taken in an environment.