Neural network: the engine of modern AI
An artificial neural network is software made of many small connected computing units that learns by strengthening or weakening those connections. It is the technology behind ChatGPT, machine translation and face recognition.
Data as of October 22, 2026
Key figures
| Level | Beginner |
|---|---|
| Origin | 1943 |
| Introduced by | Warren McCulloch et Walter Pitts |
Simple units, wired in layers
Picture a chain of tiny voters. Each one receives signals, gives them more or less weight, then passes its verdict to the next ones. On its own, a voter can do almost nothing. Stacked in successive layers, they can spot a cat in a photo, transcribe a spoken sentence or predict the next word in a text.
That is the principle of an artificial neural network. The importance each unit gives to each signal is a number, called a weight or parameter. Learning, for the network, means tuning those numbers until its answers match the correct examples: a form of machine learning.
The idea dates back to 1943, when Warren McCulloch and Walter Pitts proposed a logical model of the neuron (the 1943 paper on PubMed). It took decades, plenty of data and specialized chips such as GPUs for these networks to become truly powerful.
A concrete example
Think of reading handwritten numbers on a check or an envelope. You show the network a large number of images of digits along with the right answer. At first it guesses at random. After each mistake, its weights are nudged slightly. In the end, it can read digits written by people it has never “seen”. The same principle, at a much larger scale, lets an AI assistant draft an email for you.
Why it matters to you
Knowing that an AI runs on a neural network helps you understand its strengths and limits. It is excellent at spotting patterns in huge amounts of data. But it does not reason the way you do, it does not look things up in an internal encyclopedia, and it can produce a wrong answer with total confidence. Its mistakes often trace back to the examples it was given.
Common misconceptions
“There’s a little brain inside the computer.” No: it is arithmetic on tables of numbers, run by electronic chips.
“More neurons means a smarter AI.” Size helps, but data quality, architecture (such as the Transformer) and training matter just as much.
“Neural networks and deep learning are opposites.” Deep learning simply means neural networks with many layers.
To place these milestones in time, check the AI timeline.
Related terms
- Deep learning · Machine learning that uses neural networks with many layers; it is the technique behind today's chatbots, image and voice AI.
- Machine learning · A branch of AI where software learns patterns from examples instead of following rules written one by one by a person.
- Model parameters · The internal numbers a model tunes during training; how many it has indicates its size, not directly how good it is.
- Transformer · A neural network architecture built on attention, introduced in 2017, that underpins almost every large language model today.
- AI model · The trained “brain” behind an AI app: a large file of learned numbers that turns an input into an output, such as a reply or an image.
- GPU (graphics processing unit) · A chip built to run thousands of calculations in parallel, which became the engine for training and running AI models.
Try it
Full glossary
Frequently asked questions
Does an artificial neural network work like a brain?
It is loosely inspired by one, but it is not a copy. An artificial neuron is a simple math operation: it adds up weighted numbers and decides whether to pass a signal on. The human brain is far more complex and works differently.
How long have neural networks been around?
The first mathematical model of an artificial neuron was published in 1943 by Warren McCulloch and Walter Pitts, in the Bulletin of Mathematical Biophysics. The idea is much older than today's AI assistants.
What are the “parameters” of a neural network?
They are the numbers that set the strength of each connection. During training, the network adjusts them bit by bit to make fewer mistakes. What a model “knows” lives in those numbers, not in a database of stored sentences.
Can we tell why a neural network gave a certain answer?
Not easily. Its decisions come from the combined effect of a huge number of connections, which makes them hard to explain. Researchers work on this problem under the name interpretability.
Data sources
Data as of October 22, 2026
Report an error
Does a figure, a date or a fact look wrong? Tell us: every report is checked.