What Machine Learning Actually Means
Machine learning is the approach to building software in which the system learns to perform a task by finding patterns in data rather than by following rules that a programmer explicitly wrote. The traditional software programme for identifying spam email might have been built by writing rules: if the email contains these words, mark it as spam; if it comes from these domains, mark it as spam. The machine learning approach builds the spam filter by showing the system millions of examples of spam and non-spam email and allowing it to find the patterns that distinguish them — without the programmer specifying what those patterns should be.
The machine learning insight that makes this approach both powerful and surprising: the patterns that the system finds in large datasets are frequently better predictors than the rules that even domain experts would write explicitly. The spam filter that learns from millions of examples identifies the subtle combinations of signals that correlate with spam more reliably than any set of rules a human expert would formulate, because the patterns in large datasets reveal correlations that human experts would never think to look for. This is why machine learning has proven most valuable precisely in the domains where the patterns are too complex or too subtle for rule-based approaches to capture.
The Main Types of Machine Learning Models
The supervised learning model is the most widely deployed type: it learns from labelled examples (inputs paired with the correct outputs) and produces a model that predicts outputs for new inputs. The image classifier that learns to identify cats from millions of labelled images, the credit scoring model that learns to predict default risk from thousands of historical loan records, and the translation model that learns to convert text from one language to another from millions of parallel document pairs are all supervised learning models. The quality of the training data — its size, its accuracy, and its representativeness of the deployment context — is the primary determinant of supervised model quality.
The unsupervised learning model learns from unlabelled data, finding the patterns and structure in the data without being told what to look for. The customer segmentation model that groups customers into clusters based on their purchasing behaviour without being given predefined segment definitions, the anomaly detection model that identifies unusual network traffic patterns without being told what normal looks like, and the topic modelling system that identifies the main topics in a corpus of documents without being given a predefined topic list are all unsupervised learning models. The reinforcement learning model learns by taking actions in an environment and receiving rewards or penalties for those actions — the approach that underpins most game-playing AI and that has become increasingly important in training the large language models that power modern AI assistants.
How Models Are Trained
The training process that converts raw data and a model architecture into a working machine learning model: the optimisation algorithm iterates through the training data many times, adjusting the model’s parameters with each iteration to reduce the difference between the model’s predictions and the correct answers. This adjustment process is called gradient descent, and the rate at which parameters are adjusted is called the learning rate. The training process that runs long enough with enough data and the right architecture converges to a model whose parameters capture the statistical patterns in the training data that are relevant to the prediction task.
The training resource requirements that most shape what models can be built and by whom: the computation required to train large models has grown dramatically with the scale of the models, with the largest language models requiring thousands of specialised GPUs running for weeks or months to complete training. The compute concentration required for frontier model training has made it accessible only to a small number of well-resourced organisations, while the inference cost of running trained models has decreased rapidly through improvements in hardware efficiency and model compression techniques that make deployment more accessible.
Evaluating Model Quality
The evaluation metrics that most honestly reveal whether a machine learning model is good enough to trust for a specific application: the performance on a held-out test set (data that was not used in training, which prevents the model from memorising the training data rather than learning generalisable patterns), the performance on subgroups of the test set that represent different demographic or contextual categories (which reveals whether the model performs consistently across different populations or is systematically worse for specific groups), and the calibration (whether the model’s confidence scores accurately reflect its actual accuracy — a model that says it is 90% confident should be correct 90% of the time).
The evaluation mistake that most commonly produces models that appear excellent in testing but fail in deployment: the train-test contamination that occurs when the test set includes examples that are similar enough to training examples that the model has effectively seen them before. The model that achieves 98% accuracy on a contaminated test set may achieve only 75% accuracy on genuinely new data — a discrepancy that is not discovered until the model is in production and failing more often than expected. The evaluation rigour that prevents this requires genuine separation between training and evaluation data, ideally with time-based splits that reflect how the model will actually be used (trained on historical data, evaluated on more recent data).
When to Use Machine Learning and When Not To
The decision framework for determining whether a machine learning approach is appropriate for a specific problem: the problem should have enough labelled training data available (the minimum varies by problem type but most supervised learning applications require thousands to millions of examples to train reliably), the patterns in the data should be complex enough that rule-based approaches would be inferior (if the rules can be written clearly and would be more reliable than a learned model, the rule-based approach is usually preferable), and the application should be tolerant of errors at the model’s current accuracy level (safety-critical applications may require accuracy levels that the current state of machine learning cannot guarantee).
The machine learning application categories where rule-based or other deterministic approaches are more appropriate: compliance and regulatory decision-making where the decision logic must be fully explainable and auditable, safety-critical control systems where errors have severe physical consequences, and any application where the business must be able to explain every individual decision to the person it affects. The machine learning model that cannot explain why it made a specific prediction is unsuitable for applications where that explanation is a legal or ethical requirement — and the growth of AI regulation in most jurisdictions is expanding the set of applications where explainability is required rather than merely desirable.