Machine Learning

Rather than telling a computer exactly how to respond to every situation, machine learning allows a model to develop a mathematical representation of relationships within data. Give a system enough relevant examples and it can learn to recognise an object in an image, estimate future electricity demand, identify fraudulent transactions, recommend a product, predict whether machinery is likely to fail or determine which words are most likely to follow one another in a sentence.

Machine learning has become one of the most important technologies behind modern artificial intelligence. Recommendation systems, computer vision, speech recognition, predictive analytics and many generative AI systems are all built using machine learning techniques.

Machine learning and artificial intelligence are not, however, interchangeable terms. Machine learning is one approach within the much broader field of artificial intelligence.

What is machine learning?

Machine learning is a branch of artificial intelligence concerned with developing computational systems that can improve their ability to perform a task by learning patterns from data.

In traditional software development, a programmer typically defines the rules the computer should follow. If a certain condition occurs, the program carries out a predefined instruction. This works extremely well when the rules governing a problem can be described precisely.

Some problems are far more difficult to express through explicit rules.

Consider building software capable of identifying whether a photograph contains a dog. A programmer could attempt to describe the characteristics of a dog through thousands of instructions covering ears, fur, legs, shape, colour and size, but those rules would quickly break down when confronted with different breeds, camera angles, lighting conditions or partially obscured animals.

Machine learning approaches the problem differently. Instead of attempting to describe every characteristic manually, a model can be trained using examples. During training, the system adjusts its internal parameters so that patterns within the examples become mathematically associated with the correct outcomes.

Once trained, the model can use what it has learned to make predictions about data it has not previously encountered.

This ability to generalise from examples is central to machine learning.

What does ML stand for?

ML stands for machine learning.

The term refers to computational methods through which a system develops or adjusts its behaviour using data rather than relying entirely on instructions written explicitly by a programmer.

The word "learning" can sometimes create confusion because computers do not necessarily learn in the same way humans do. A machine learning system does not need to understand the meaning of the information it processes. Learning usually describes a mathematical optimisation process in which the parameters of a model are adjusted to improve its performance against a particular objective.

A model trained to distinguish cats from dogs, for example, is not necessarily developing a human-like concept of what a cat or dog is. It is identifying statistical relationships within the data that allow it to separate the two categories reliably.

Machine learning can therefore produce highly capable behaviour without requiring the system to experience learning in the biological or conscious sense associated with people.

Who invented machine learning?

Machine learning was not invented by a single person. It developed from research spanning artificial intelligence, statistics, mathematics, computer science, information theory and neuroscience.

One of the researchers most closely associated with the term is American computer scientist Arthur Samuel.

During the 1950s, Samuel developed computer programs capable of playing checkers and improving their performance through experience. His work was significant because the program did not rely solely on a complete set of strategies manually specified by its creator. It could use previous games to improve the way it evaluated possible moves.

Samuel is commonly credited with popularising the term "machine learning" during this period.

Other foundational contributions came from researchers working on neural networks, pattern recognition, statistics and early artificial intelligence. Frank Rosenblatt's development of the perceptron during the 1950s helped establish an important early model for artificial neural networks, while later advances in optimisation, computing power and data availability allowed machine learning methods to become dramatically more capable.

Machine learning subsequently developed through several overlapping traditions rather than following a single technological path. Statistical learning, decision trees, support vector machines, ensemble methods and neural networks all contributed to the field that exists today.

How does machine learning work?

Machine learning works by using data to adjust a mathematical model so that it becomes better at performing a specified task.

The exact process varies considerably according to the type of machine learning being used, but most systems involve several common stages.

First, a problem must be defined. A business might want to predict customer demand, a bank might want to identify fraudulent transactions, or a computer vision system might need to determine what objects appear within an image.

Relevant data is then collected and prepared. This data represents examples from which the system can learn.

A machine learning algorithm is used to train a model on that data. During training, the model produces outputs, compares those outputs against an objective and adjusts its internal parameters in an attempt to reduce error or improve performance.

This process may be repeated thousands, millions or even billions of times depending on the scale and complexity of the system.

Once training has finished, the resulting model can be evaluated using data it did not encounter during training. This helps determine whether the model has learned patterns that generalise beyond the original examples rather than simply memorising its training data.

If the model performs sufficiently well, it can be deployed and used to make predictions or decisions when new inputs arrive.

The crucial distinction is therefore between the algorithm used to learn and the model produced through that learning process.

What is a machine learning model?

A machine learning model is a mathematical system that has been adjusted through training to identify relationships within data and produce an output from new inputs.

The model is the result of the learning process rather than simply the computer program performing it.

A model might receive information about a house and predict its likely sale price. Another might analyse an email and estimate whether it is spam. A computer vision model might receive an image and assign probabilities to the objects it believes the image contains.

Modern machine learning models can range from relatively simple mathematical structures containing a small number of variables to enormous neural networks containing billions or even trillions of adjustable parameters.

The size of a model does not automatically determine how useful it is. The appropriate model depends on the problem, available data, required accuracy, computing resources, interpretability requirements and the environment in which the system will operate.

In many business and industrial applications, a specialised model designed for a clearly defined problem may be more useful than a much larger general-purpose AI model.

What is a machine learning algorithm?

A machine learning algorithm is a method used to identify patterns within data and adjust a model during the learning process.

Different algorithms make different assumptions about how information should be represented and how useful relationships should be discovered.

Examples include linear regression, logistic regression, decision trees, random forests, gradient boosting, support vector machines, clustering algorithms and neural networks.

Some algorithms are particularly effective for structured numerical data. Others are designed for images, language, time series or extremely high-dimensional information.

Choosing an algorithm is therefore only one part of developing a machine learning system. The quality and relevance of the data, the way the problem is formulated, the objective used during training and the conditions encountered after deployment can all have a substantial effect on performance.

What are the main types of machine learning?

Machine learning is commonly divided into several categories according to the kind of information available during training and the way the system receives feedback.

The most frequently discussed categories are supervised learning, unsupervised learning and reinforcement learning. Semi-supervised and self-supervised learning are also important, particularly within modern AI.

Supervised learning

Supervised learning trains a model using examples for which the correct outcome is already known.

The training data therefore contains both inputs and labels.

A model trained to recognise spam emails might receive thousands of emails that have already been labelled as either spam or legitimate. The system attempts to identify relationships between characteristics of the emails and the correct classifications.

Supervised learning is widely used for classification and regression problems.

Classification predicts a category, such as whether a transaction is fraudulent or whether an image contains a particular object. Regression predicts a numerical value, such as tomorrow's energy demand, the likely price of a property or the expected lifetime of an industrial component.

Unsupervised learning

Unsupervised learning works with data that does not contain predetermined labels describing the correct answer.

Instead, the system attempts to identify structures, similarities or relationships within the data itself.

Clustering is a common example. A machine learning system might analyse customer behaviour and discover groups of customers with similar purchasing patterns without being told beforehand what those groups should be.

Unsupervised learning can be useful for discovering patterns in large datasets, reducing dimensionality, identifying unusual behaviour and exploring information where labelled examples are unavailable.

Semi-supervised learning

Semi-supervised learning combines a relatively small amount of labelled data with a larger quantity of unlabelled data.

Producing high-quality labels can be expensive or time-consuming, particularly when specialist expertise is required. A medical dataset, for example, may contain millions of images while only a limited number have been reviewed and labelled by clinicians.

Semi-supervised methods attempt to use information contained within the additional unlabelled data to improve performance.

Self-supervised learning

Self-supervised learning allows training signals to be generated from the structure of the data itself rather than requiring humans to provide a label for every example.

This approach has become particularly important in modern generative AI.

Large language models can, for example, learn from enormous collections of text by repeatedly attempting to predict missing or subsequent elements within sequences. The relationships learned through this process can result in representations of language that are useful across many different tasks.

Self-supervised learning has helped make it possible to train highly capable models using the enormous quantities of text, images, audio and video available digitally without requiring every individual example to be manually labelled.

Reinforcement learning

Reinforcement learning trains systems through interaction and feedback.

Instead of learning primarily from labelled examples, an agent performs actions within an environment and receives signals indicating how desirable the resulting outcomes are.

The objective is to learn a strategy, often called a policy, that maximises cumulative reward.

Reinforcement learning has been used in areas including robotics, game-playing, resource allocation and AI model alignment. It differs substantially from conventional supervised learning because the system must consider how its actions influence future states as well as immediate outcomes.

What is the difference between machine learning and artificial intelligence?

Artificial intelligence is the broader field concerned with creating computational systems capable of behaviours associated with intelligence. Machine learning is one set of methods used to achieve that objective.

The relationship is often described as machine learning being a subset of artificial intelligence.

Not every AI system uses machine learning. Earlier artificial intelligence systems frequently relied on rules, symbolic reasoning, logic, search or expert knowledge encoded directly by programmers.

Machine learning instead allows aspects of a system's behaviour to emerge through exposure to data.

The distinction matters because the extraordinary success of machine learning over recent decades has caused the terms AI and machine learning to become closely associated. Much of the artificial intelligence people encounter today does use machine learning, but the two concepts remain technically different.

Artificial intelligence describes the broader ambition. Machine learning describes one important family of approaches used to pursue it.

What is the difference between machine learning and deep learning?

Deep learning is a specialised area within machine learning based on artificial neural networks containing multiple processing layers.

Machine learning is therefore the broader category.

Traditional machine learning approaches may depend heavily on humans deciding which characteristics of the data should be provided to the model. These characteristics are often called features.

A system predicting house prices, for example, might be given features including floor area, location, number of bedrooms and age of the property.

Deep neural networks can increasingly learn useful representations directly from more complex raw data. A computer vision system can learn visual features from pixels, while a language model can develop representations of linguistic relationships from sequences of text.

Deep learning has been particularly important to advances in image recognition, speech recognition, natural language processing, generative AI and large language models.

Its success has been supported by access to enormous datasets, specialised computing hardware and improvements in neural-network architectures and training methods.

Is generative AI machine learning?

Most modern generative AI systems are built using machine learning, particularly deep learning.

Generative AI refers to systems designed to create new outputs such as text, images, audio, software or video.

Large language models, image-generation systems and many other contemporary generative models are neural networks trained using large quantities of data.

Generative AI and machine learning should not therefore be treated as interchangeable terms.

Machine learning also powers systems that generate nothing at all. Fraud detection, demand forecasting, recommendation systems, anomaly detection, predictive maintenance and medical classification can all use machine learning without being forms of generative AI.

Generative AI is one increasingly prominent application of machine learning rather than a replacement for the wider field.

What is training in machine learning?

Training is the process through which a machine learning model adjusts its internal parameters using data.

At the beginning of training, a model may perform poorly because its parameters have not yet been adjusted to represent useful relationships.

The model processes examples, produces outputs and receives information about how well those outputs satisfy the training objective. An optimisation process then changes the parameters in a direction intended to improve performance.

Repeated over many examples and training cycles, these adjustments allow the model to develop mathematical relationships that can later be applied to new inputs.

Training should be distinguished from inference.

Training is the process through which the model learns its parameters. Inference is the process of using the trained model to generate a prediction, classification or other output.

A model may require enormous computing resources to train but considerably less computing power each time it performs inference.

Does machine learning continue learning after deployment?

Not necessarily.

One of the most common misconceptions surrounding machine learning is that a system described as "learning" must continuously improve whenever it receives new information.

Many machine learning models do not operate this way.

A conventional model is often trained using a historical dataset, evaluated and then deployed with its parameters effectively fixed. New data is processed during inference, but the model itself does not automatically learn from every new example it encounters.

If conditions change, the organisation operating the model may need to collect additional data and retrain or update it.

This distinction becomes particularly important in environments that change over time. Customer behaviour can shift, financial markets can evolve, machines can wear, supply chains can change and physical systems can enter operating conditions that were poorly represented in historical training data.

When the statistical relationships encountered after deployment differ from those present during training, model performance can deteriorate. This phenomenon is closely associated with model drift.

Continual learning addresses a related challenge by investigating how machine learning systems can safely incorporate new information and adapt over time without repeatedly starting the learning process from scratch or catastrophically forgetting useful knowledge acquired previously.

What is overfitting in machine learning?

Overfitting occurs when a machine learning model becomes too closely adapted to its training data and consequently performs poorly on new information.

A model should ideally learn patterns that generalise beyond the examples it has already seen.

If it instead learns irrelevant details, noise or accidental relationships contained within the training dataset, its training performance may appear excellent while its real-world performance is significantly worse.

The opposite problem, underfitting, occurs when a model is too simple or insufficiently trained to capture important relationships even within the available data.

Machine learning development therefore involves finding a balance between learning enough complexity to solve the problem and retaining the ability to generalise to unfamiliar examples.

Evaluation using separate validation and test data is one of the methods used to assess whether a model has achieved this balance.

What is machine learning used for?

Machine learning is used across an enormous range of consumer, scientific, industrial and commercial applications.

Search engines use machine learning to understand queries and rank information. Streaming services and online retailers use it to predict which films, music or products an individual might prefer. Banks use machine learning to identify unusual transactions and estimate financial risk.

Manufacturers can use machine learning to detect product defects, monitor equipment and predict failures before they occur. Logistics organisations use it to forecast demand and improve routing. Energy companies use machine learning to forecast electricity consumption, generation and market conditions.

Healthcare applications include medical image analysis, disease-risk prediction and biological research. Cybersecurity systems can use machine learning to identify patterns associated with malicious behaviour, while computer vision models allow machines to interpret images and video.

Machine learning also sits beneath many of the generative AI applications that have made artificial intelligence far more visible to the general public.

The result is that people interact with machine learning constantly, even when no product is explicitly labelled as an AI or ML system.

Why is machine learning important?

Machine learning allows computers to address problems for which writing every relevant rule manually would be impractical or impossible.

Real-world environments frequently contain enormous numbers of variables and relationships. Some can be described mathematically by people, while others become apparent only after analysing large quantities of data.

Machine learning provides methods for extracting useful predictive relationships from that information.

This has changed the role of software. Traditional computer programs primarily execute instructions that programmers already know how to specify. Machine learning systems can instead help discover relationships that developers did not explicitly encode.

The significance of this shift extends beyond automation.

A system can estimate future events, classify unfamiliar information, identify patterns across datasets too large for humans to inspect manually and adapt its behaviour according to relationships discovered through experience.

This ability is one of the foundations on which much of modern artificial intelligence has been built.

What are the limitations of machine learning?

Machine learning systems are only as useful as the conditions under which their predictions remain reliable.

Poor-quality, incomplete or biased data can produce misleading models. A system may discover correlations that appear predictive without representing meaningful causal relationships. Training data may also underrepresent particular situations or groups, resulting in unreliable performance when those cases appear after deployment.

Machine learning models can struggle when the environment changes. A model trained using historical patterns may gradually become less accurate if the underlying relationships that produced those patterns no longer hold.

Complex models can also be difficult to interpret. It may be possible to measure that a neural network performs a task accurately without being able to provide a simple explanation for every individual prediction it makes.

Computational cost can present another limitation. Training extremely large models may require substantial quantities of specialist hardware, electricity and data.

Machine learning also does not automatically determine whether a prediction should be acted upon. Questions surrounding accountability, fairness, safety, privacy and the consequences of incorrect decisions remain human and organisational concerns even when the prediction itself is produced by a machine.

Understanding machine learning therefore requires understanding not only what a model can predict, but where its assumptions may fail.

What is the future of machine learning?

Machine learning is moving towards systems that are more capable, more efficient and increasingly able to operate across different types of information and environments.

One major direction involves general-purpose models trained on large quantities of text, images, audio, video and other data. These models can develop representations that are useful across many tasks rather than being built exclusively for a single narrow application.

Another important direction concerns efficiency. Researchers and technology companies are investigating how increasingly capable machine learning can be delivered using less computing power, less energy and smaller quantities of labelled data.

Adaptation is likely to become increasingly important as AI moves into dynamic real-world environments.

Most conventional machine learning still separates training from deployment. The model learns first and is then used. As AI becomes embedded in robotics, industrial systems, infrastructure, financial markets and other changing environments, there is growing interest in systems capable of learning safely from new experience after deployment.

Continual learning, online learning and related approaches explore how models might update their knowledge while remaining stable, reliable and resistant to forgetting previously learned information.

Machine learning is therefore evolving beyond the question of how a model can learn from a dataset. An increasingly important question is how intelligent systems can continue learning as the world around them changes.