supervised learning

basicwidespread

A machine learning method in which the machine learns from examples with correct answers how to give a good answer on its own to new cases.

In plain terms

Imagine teaching a child about animals from a picture book: you point to the image and say 'this is a dog', 'this is a cat'. After a while, the child recognizes a dog they have never seen before. Supervised learning does the same with machines: we write the correct solution next to every example, which is called a label. However, the machine typically needs far more examples than a child, often in the thousands or millions.

The word 'supervised' does not mean someone sits next to the machine continuously. It means the correct answer is already present in the training material, and the machine measures its own guesses against it.

It is good for two main tasks. One is classification: spam or not spam, benign or malignant lesion. The other is numerical estimation: how much a flat is worth, how much electricity consumption will be tomorrow.

It differs from other learning methods in that someone must first provide the correct answer. In unsupervised learning, the machine searches for patterns without labels; in reinforcement learning, it learns through trial and error based on rewards or penalties.

How it works

They collect examples and assign the correct answer to each: for instance, thousands of emails, each labeled "spam" or "not spam." Labeling is mostly done by humans.
Examples are split into two parts: the model learns from the larger portion, while the smaller part is set aside for testing.
The model makes a guess for each training example, the guess is compared with the correct answer, and the model's internal settings are adjusted a little at a time based on the error. This is repeated many thousands of times.
The model's accuracy is measured on held-out, previously unseen examples. This reveals whether it truly learned the underlying rules or merely memorized the training material.
If the result is good enough, the model starts being used live on new data and is occasionally retrained with fresh examples because the world changes.

An everyday example

When your email automatically moves a suspicious prize notification to the spam folder, a supervised learning model is at work. It learned from countless emails that humans previously marked as spam. When you mark an email as spam, you provide another labeled example with which the provider can further train the filter.

Who is affected

everyoneBehind spam filters, your phone's face recognition, and translation programs, there are largely models trained in this way.

bank customer, investorCredit assessment and fraud detection typically rely on models trained on past cases with known outcomes, so the decision about you depends on the quality of historical data. → Money

patient, doctorSystems analyzing X-ray and dermatological images learn from doctor-labeled pictures, so they are only as reliable as the labels. → Health

entrepreneur, executiveIf you want to make predictions from your own data (churn, demand, defects), the biggest cost is usually producing high-quality, labeled data.

A common misconception

MythIn supervised learning, a human oversees and verifies every decision made by the machine.

In realitySupervision applies only to the training period and means that the correct answer is present alongside the examples. In live use, the model decides independently unless human oversight is specifically built in.

Limits and risks

It is only as good as the labels: if the training data is incorrect or biased, the model learns and repeats this faithfully.
Labeling is expensive and slow, especially where experts are required, such as for medical images.
In situations unlike the training examples, the model can confidently err without signaling this itself.
Sometimes the model memorizes the training material instead of learning general rules (overfitting), so accuracy must always be measured on unseen data.

Similar, but not the same

unsupervised learning – There are no correct answers next to the examples; the machine itself searches for groups and patterns in the data.
reinforcement learning – There, the machine does not learn from ready-made solutions but tries out actions, and its behavior is shaped by rewards or penalties received afterward.
self-supervised learning – There, the label is not given by a human but by the data itself: for example, the model tries to guess the next word in a text. This is how pretraining of large language models works.

Related concepts

Broader

machine learning, artificial intelligence

Narrower

classification, regression

On the edge

unsupervised learning, reinforcement learning, self-supervised learning, fine-tuning, training data, data labeling

A brief history

One of the first machine learning models to learn from labeled examples was Frank Rosenblatt's perceptron, described in a paper published in 1958. Supervised deep neural networks gained major momentum in 2012, when a network called AlexNet, trained on more than one million labeled images, won the ImageNet image recognition competition by a wide margin.

Where things stand

By the end of 2026, supervised learning is one of the oldest and most widely used methods in machine learning, from spam filtering to medical image analysis. Large language models are based on self-supervised pre-training. They are then trained for instruction following via supervised fine-tuning, i.e., with example question-answer pairs, typically supplemented with reinforcement methods.

Status: 2026 November

Also known as

supervised learning · supervised machine learning · controlled learning · supervised learning · supervised learning · supervised learning · supervised machine learning · supervised ML

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.