Machine learning

Machine learning is a field of artificial intelligence in which computer programs improve their performance on a task by learning from data, rather than being told exactly what to do step by step. Instead of a programmer writing rules for every situation, a model is shown examples and adjusts itself to capture the underlying patterns.

This approach has made it possible to solve problems that are easy for people but hard to describe as rules — recognising a face, understanding speech, or predicting which customers might leave a service.

The core idea

At its heart, machine learning is about finding patterns that generalise. A model is trained on data, producing predictions or decisions; a measure of error tells the model how wrong it is; and an algorithm adjusts the model's internal parameters to reduce that error. Repeated over many examples, this process gradually improves performance.

The goal is not to memorise the training data but to perform well on new, unseen data. A model that fits its training examples too closely but fails on fresh data is said to overfit; one that is too simple to capture the pattern underfits.

Main types of learning

Supervised learning. The model learns from labelled examples — inputs paired with the correct answers — and then predicts labels for new inputs. Common tasks include classification (choosing a category) and regression (predicting a number).

Unsupervised learning. There are no labels. The model looks for structure in the data, such as grouping similar items (clustering) or reducing many variables to a few meaningful ones.

Reinforcement learning. An agent learns by interacting with an environment, taking actions and receiving rewards or penalties. Over time it learns a strategy that maximises reward. This approach suits problems like game playing and control.

Semi-supervised and self-supervised methods blend these ideas, using large amounts of unlabelled data together with smaller labelled sets.

Common algorithms

Classical methods include linear and logistic regression, decision trees, random forests, support vector machines, and k-nearest neighbours. These remain effective for many tasks and are often easier to interpret. Neural networks, especially deep networks with many layers, excel at complex data such as images, audio, and text. Deep learning has driven major advances in recent years, powered by large datasets and specialised hardware.

From data to deployment

A typical project begins by defining the problem and gathering data, which must then be cleaned and prepared. Data is split into training, validation, and test sets so performance can be judged fairly. After training and tuning, the model is evaluated and, if good enough, deployed in a real system, where it may be monitored and retrained as conditions change.

Data quality matters enormously. Biased, incomplete, or unrepresentative data leads to unreliable models, no matter how sophisticated the algorithm.

Applications

Machine learning underpins recommendation systems, spam filters, fraud detection, medical image analysis, speech recognition, machine translation, and predictive maintenance, among many others. It also supports scientific work, from classifying galaxies to predicting protein structures. In industry it helps with forecasting, pricing, and quality control.

Challenges and limits

Machine learning is powerful but not magic. It demands good data and careful evaluation, and can fail in unexpected ways when confronted with situations unlike its training examples. Models can be hard to interpret, which is a problem in settings where explanations are required. They can encode societal biases, raise privacy concerns, and be vulnerable to manipulation. Building, running, and maintaining them also costs energy and expertise.

Features and evaluation

How data is represented matters as much as the algorithm. In classical machine learning, engineers often craft "features" — measurable properties of the data — that make patterns easier to detect. Deep learning reduces this manual work by learning useful representations directly from raw inputs such as pixels or words. Once a model is trained, it is evaluated on data it has not seen, using metrics suited to the task, such as accuracy, precision, or error rates. Cross-validation and separate test sets help ensure the results reflect genuine performance rather than luck.

Regularisation, careful tuning, and enough varied data all help a model generalise instead of memorising.

Ethics and responsible use

Because machine learning shapes decisions about people — loans, hiring, medical care — fairness and accountability matter. Models can absorb historical biases in their training data and make them seem objective. Keeping humans in the loop for high-stakes choices, documenting how models were built, testing for disparate impact, and protecting personal data are common safeguards. Transparency is another goal: some models can be explained, while complex ones are harder to interpret, creating a trade-off between accuracy and clarity.

Common misconceptions

One misconception is that more data always means better results; relevance and quality matter more than sheer volume. Another is that a model "understands" its task the way a person does — it detects statistical patterns. Some assume machine learning needs no human involvement, when in fact people define goals, curate data, and judge acceptable performance. Finally, high accuracy on a test set does not guarantee success in the messy real world.

Key points