Machine Learning

Traditional programming works by writing rules. You analyse a problem, break it into logical steps, and translate those steps into code. The program does exactly what you tell it, nothing more. This works well for tasks with clear, consistent rules - calculating a payroll, sorting a list, simulating physics. It fails for tasks where the rules are too complex to write down, or where they change depending on context.

Machine learning flips this. Instead of writing rules, you show the computer many examples of the problem and its correct answer, and the computer figures out the rules itself. Show a machine learning system ten thousand emails labelled "spam" and "not spam," and it learns to classify new emails without being told what patterns to look for. It discovers them itself.

The core mechanism is an objective function - a measure of how wrong the model's predictions are. During training, the model makes predictions, its error is measured, and its internal parameters are adjusted slightly to reduce that error. Repeat this millions of times across millions of examples and the model's predictions steadily improve. This process is called gradient descent.

Machine learning has several major subfields. Supervised learning is the most common: you provide labelled examples and the model learns to map inputs to outputs. Unsupervised learning finds patterns in data without labels. Reinforcement learning trains agents to make sequences of decisions by rewarding good outcomes. Most of what we call AI today is supervised or self-supervised learning applied at enormous scale.

The critical insight is that machine learning models do not just memorise their training data. They generalise: a model trained on labelled cat photos can identify a cat it has never seen before. This ability to generalise to new situations is what makes machine learning useful, and understanding its limits - where generalisation breaks down - is essential for using these systems reliably.