Definition
Machine learning is the subfield of artificial intelligence concerned with building systems that improve at a task through exposure to data rather than through explicit rule-writing. Rather than a programmer specifying every step of a solution, a machine learning system infers the relevant patterns, relationships, or decision boundaries from examples, and then applies what it has learned to new, unseen inputs.
The central idea is that many real-world problems are too complex, variable, or poorly understood for a human to enumerate all the rules that govern them. Recognising faces, filtering spam, forecasting demand, and diagnosing images from medical scans all involve patterns that are easier to learn from thousands of examples than to specify by hand. Machine learning automates the extraction of those patterns.
Three broad paradigms organise the field. In supervised learning, a system is trained on labelled examples — input–output pairs — and learns a mapping between them, then predicts outputs for new inputs. In unsupervised learning, the system receives only inputs and discovers structure (clusters, latent dimensions, generative factors) without labels. In reinforcement learning, an agent takes actions in an environment, receives reward or penalty signals, and adjusts its behaviour to maximise cumulative reward over time. Each paradigm suits different problem types and data availability situations.
Why it matters
How it works
From data to model
A machine learning pipeline begins with data collection and cleaning, followed by feature engineering (deciding which aspects of the input to represent and how). The model — a parameterised mathematical function — is then fitted to the data by optimising a loss function that measures how far predictions are from targets. Gradient-based optimisation, particularly stochastic gradient descent, adjusts the model's parameters iteratively to reduce this loss. After training, the model is evaluated on a held-out test set to estimate its generalisation performance.
The choice of model architecture encodes assumptions about the problem structure. Linear models assume the output is a linear combination of inputs; decision trees partition the input space with axis-aligned splits; neural networks compose many layers of non-linear transformations that can represent highly complex functions. The tradeoff between model expressiveness and the risk of overfitting runs through every architectural decision.
Evaluation and deployment
Choosing the right evaluation metric matters as much as choosing the right model. Accuracy is misleading when classes are imbalanced; in fraud detection, for instance, recall (catching real fraud) may matter far more than precision. Metrics like the area under the ROC curve, F1 score, and mean squared error each highlight different aspects of model performance. Hyperparameter tuning — adjusting settings like learning rate, regularisation strength, and network depth — is usually performed on a separate validation set to avoid contaminating the test set.
Once deployed, models encounter distribution shift: the real world changes over time, and the patterns in production data may differ from those in training data. Monitoring, periodic retraining, and feedback loops are essential parts of any mature machine learning system, not one-time engineering concerns.
Where it goes next
Machine learning is the engine of most current AI progress: large language models, image generation, protein structure prediction, and game-playing agents all rest on the same fundamental learning-from-data paradigm, scaled up enormously in data and compute. The next frontiers include more sample-efficient learning (learning from fewer examples), better uncertainty quantification (knowing what the model does not know), and causal reasoning (understanding mechanisms, not just correlations). As ML systems become more deeply embedded in infrastructure, the problems of fairness, robustness, and accountability will grow in importance alongside the technical ones.