Neural network

The name comes from a loose analogy with brain cells, but a neural network is mathematics, not biology. Each unit takes some numbers, weighs them, adds them up and passes a result to the next layer. The weights are the parameters that training adjusts.

Neural networks are one family of machine learning models. "Deep learning" simply means neural networks with many layers, and almost every headline AI system, including large language models, is a deep neural network. The design itself is decades old; what made it powerful is the scale of data and computing power available to train it.

How it works

Data enters at the input layer as numbers: pixel colours, or words converted into numeric form. Each hidden layer transforms those numbers, picking out increasingly abstract features. In an image model, early layers might respond to edges, later ones to shapes, and the final layers to whole objects. The output layer gives the answer, such as a label or the next word.

Training compares the output with the correct answer and works backwards through the layers, adjusting every weight slightly to reduce the error. Repeating this millions of times is what makes the network accurate. Once trained, a neural network is fixed. It does not learn from each new photo or question unless its makers deliberately train it again. This explains a common surprise: correcting an AI tool in one conversation rarely changes how it behaves for the next person. Any improvement has to come from new training, from fresh instructions in the prompt, or from documents the tool looks up while answering.

Different tasks use different arrangements of layers. Image tasks long relied on convolutional networks, which scan a picture in small patches. Language tasks moved to the transformer, which compares every word with every other word. The arrangement matters less to a business user than the training: two networks with the same design can behave very differently depending on the data and the effort spent tuning them.

Example

A photo app that blurs the background behind a person uses a neural network trained on many images where people were outlined by hand. Given a new photo, the network marks which pixels belong to the person, and the app blurs the rest.

The weakness shows at the edges. A person holding a plant with leaves that overlap their hair may get a patchy outline, because few training images looked like that. The network does not know it is unsure; it simply produces its best guess, which is why such tools usually let you correct the outline by hand. The app maker improves it by collecting examples of these failures and adding them to the next round of training.

Why it matters for your business

You will rarely build a neural network yourself, but understanding the idea explains AI behaviour you will meet: models are strong on patterns similar to their training data, weak on unusual cases, and hard to explain because their knowledge is spread across billions of weights rather than stored as readable rules. It also explains why a tool can handle ninety-nine ordinary cases and then fail oddly on the hundredth, so any process that relies on one should include a simple way for a person to spot and fix those failures before customers see them.

Common mistakes

• Believing a neural network reasons like a human brain because of the name.
• Expecting a clear explanation of why a neural network made a particular decision.
• Assuming a bigger network is always better, when a smaller one trained on the right data can do a narrow job well.

Questions people ask

What is the difference between a neural network and deep learning?

Deep learning is the use of neural networks with many layers. A small neural network with one or two layers is still a neural network but not usually called deep learning. Nearly all modern AI tools for language, images and speech use deep networks.

Are large language models neural networks?

Yes. A large language model is a very large neural network, usually built on the transformer design, trained on huge amounts of text to predict the next token. Its billions of parameters are the connection weights inside that network.

Why are neural networks called black boxes?

Because their knowledge is spread across millions or billions of numeric weights rather than written as rules a person can read. You can see what goes in and what comes out, but tracing exactly why a given answer was produced is difficult, though researchers are building tools to interpret them.