What is a neural network?

Almost every impressive AI system today — from face unlock to chatbots — is built on a neural network. It sounds futuristic, but the core idea is surprisingly simple.

What is a neural network, exactly?

A neural network is a computing system made of many simple units called neurons, connected together in layers. It’s loosely inspired by the way brain cells link up, though it works very differently from real biology.

Each artificial neuron does something basic: it receives some numbers, weighs them up, adds them together, and passes on a result. On its own, one neuron isn’t clever. But connect thousands or millions of them, and the network as a whole can recognise faces, understand speech, or write text.

Neural networks are the foundation of deep learning, which is itself the most powerful branch of machine learning.

How is a neural network structured?

A typical neural network has three kinds of layers:

• Input layer — where the data enters (for example, the pixels of an image).
• Hidden layers — the middle layers that do the real work of transforming the data. There can be one, or hundreds.
• Output layer — where the answer comes out (for example, “this is a cat” with 97% confidence).

Information flows from input, through the hidden layers, to the output. Each connection between neurons has a weight — a number that controls how much influence one neuron has on the next.

How does a neural network learn?

Learning a neural network means finding the right weights. Here’s the loop, in plain terms:

1. Guess. The network takes an example and produces an output.
2. Check. It compares its output to the correct answer and measures the error.
3. Adjust. It nudges its weights slightly to reduce that error — a process called backpropagation.
4. Repeat. It does this across huge numbers of examples until the predictions get reliably good.

This is the same training idea described in how AI works: make a prediction, measure the mistake, adjust, and repeat. Over millions of rounds, the weights settle into values that capture real patterns in the data.

Why are neural networks so powerful?

The strength of neural networks is that they learn their own features. You don’t have to tell the network what an “ear” or a “verb” is — given enough examples, it discovers the useful patterns by itself, layer by layer.

This flexibility is why the same basic idea can be applied to images, audio, language, and more. The large language models behind modern chatbots are giant neural networks of a special design called transformers, trained on enormous amounts of text.

What are the limits of neural networks?

Neural networks are remarkable but imperfect:

• They need lots of data. Too few examples and they fail to generalise.
• They’re hard to interpret. With millions of weights, it’s tough to explain a specific decision.
• They can be confidently wrong. A network may give a wrong answer with high certainty if it meets something unlike its training data.

Are there different kinds of neural networks?

Yes. The basic structure stays the same, but networks are arranged differently depending on the job:

• Feedforward networks send information straight from input to output. They’re the simplest type and a good starting point for understanding the idea.
• Convolutional networks (CNNs) are tuned for images, scanning a picture in small patches to detect visual features.
• Recurrent networks (RNNs) handle sequences like speech or text by carrying information forward from one step to the next.
• Transformers are the modern design behind language AI. They can consider all parts of an input at once, which is why they power today’s chatbots.

All of these are still neural networks — collections of weighted connections that learn from data — just wired up in ways that suit different problems.

How do neural networks relate to the brain?

The “neural” name comes from a loose analogy with biological brain cells, which also pass signals through connections. But the comparison only goes so far. A real brain has around 86 billion neurons with staggeringly complex chemistry, learns continuously, and uses very little energy. Artificial networks are simplified mathematical models that need far more data and power to learn far narrower skills. The brain inspired the idea; it isn’t a blueprint.

A neural network in one sentence

A neural network is a web of simple connected units that learns by tuning the strength of its connections — a simple idea that, scaled up massively, underpins almost all of modern artificial intelligence.