Neural network (NN)

basicwidespread

A neural network is a computer model that learns to recognize patterns from many examples, loosely inspired by the network of neurons in the human brain.

In plain terms

In traditional programming, humans write the rules step by step. For a neural network, you do not give rules but examples: thousands of cat images, millions of sentences, from which it filters out what they have in common.

Imagine a huge mixing console with millions, even billions, of tiny knobs. During learning, after each attempt, the machine turns the knobs slightly in the direction that reduces the error, repeating this until the output is good enough. 'Knowledge' is not in a sentence or rule, but in the position of the knobs.

This makes it good for tasks that are difficult to codify into rules: face and voice recognition, translation, text writing, and image generation. Today's chatbots, image generators, and AI agents are all built on neural networks.

How it works

The network consists of small computing units, artificial neurons, arranged in layers: an input layer, one or more intermediate (hidden) layers, and an output layer.
Every connection has a weight, meaning a number that says how much a given signal matters. The input (for example, pixels of an image) passes through layers as numbers. Every neuron adds up the incoming signals multiplied by their weights, then passes the result through a simple function. It is thanks to this function that the network can learn even complex relationships.
At the end, the network gives an answer, for example, "87 percent chance it is a cat". During training, this is compared with the correct answer, and the error is calculated.
Then the error is propagated backward through the network, and it is calculated how much each weight is responsible for it. This is called backpropagation. Based on this, every weight is adjusted slightly in the direction that reduces the error.
This is repeated on countless examples; once the network is ready, it only needs to be used: it responds to new input with the learned weights.

An everyday example

When your phone unlocks the screen with facial recognition, or you search for "dog" in your photo album and receive all your dog photos, a neural network is at work. The same happens when your email filters spam or a chatbot answers your question.

Who is affected

everyoneNeural networks operate in your phone, search engine, translator, and chatbots, even when you do not see them.

employeeAI tools appearing in workplaces are almost all built on this, so it is worth knowing why they sometimes err confidently. → Work & HR

patient, doctorNeural networks increasingly assist in analyzing X-ray, CT, and dermatological images, but the responsibility for the decision remains with the doctor. → Health

developerIf you integrate AI into a product, the network's architecture, size, and training data determine what to expect from it. → Code

A common misconception

MythA neural network works like the human brain, so it thinks and understands.

In realityIt only borrowed the idea from the brain: in reality, it is a long series of operations performed with numbers that learns statistical patterns. Whether this can be called understanding is a topic of debate among researchers.

Limits and risks

It is hard to look inside: often even the creators cannot say exactly why a certain response was given (this is the "black box" problem).
It is only as good as its training data: if the data is incomplete or biased, the network's output will be skewed too.
It can confidently be wrong, especially in situations unlike those it was trained on.
Training large networks requires vast amounts of data, expensive chips, and lots of electricity.

Similar, but not the same

artificial intelligence – Artificial intelligence is the broader field, while the neural network is one of its methods – currently the most widespread.
deep learning – Deep learning refers to training multi-layered neural networks, meaning a specific, large-scale method of using neural networks.
large language model – A large language model is a specific type of neural network trained on text; not every neural network is a language model.

Related concepts

Broader

artificial intelligence, machine learning

Narrower

deep learning, transformer, convolutional neural network, large language model

On the edge

training, parameter, training data, backpropagation, algorithm

A brief history

The first mathematical neuron model was described by Warren McCulloch and Walter Pitts in 1943. The first learning-capable network, the perceptron, was presented by Frank Rosenblatt in 1957–58. Training multi-layer networks became practically feasible with the 1986 popularization of backpropagation (David Rumelhart, Geoffrey Hinton, and Ronald Williams). The real breakthrough came in the 2010s with abundant data and powerful graphics chips. In 2024, John Hopfield and Geoffrey Hinton received the Nobel Prize in Physics for their work underpinning neural networks.

Where things stand

At the end of 2026, virtually every known AI service is built on neural networks, from phone face recognition and chatbots to AI agents. Research focuses mainly on making these networks cheaper, more reliable, and more transparent.

Status: 2026 November

Also known as

neural network · artificial neural network · neural network · neural network · neural network · artificial neural network · ANN · neural net

Sources

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