What is a neural network?
A neural network is a computer model that learns patterns from data instead of following rules a programmer wrote by hand. It is built from many small units called neurons, arranged in layers. Each connection between neurons carries a weight, a number that decides how much one neuron influences the next.
Information flows through the layers. The first layer takes the input, for example the pixels of a photo. The middle layers (often called hidden layers) combine and reshape that signal. The last layer produces the answer, such as "this is a cat". During training, the network adjusts its weights again and again until its answers match the examples it has seen.
The "deep" in deep learning simply means a network with many hidden layers. More layers let the network capture more complex patterns, which is why neural networks now sit behind image recognition, voice assistants, and large language models.
In plain words
Picture a huge team of people passing notes. Each person looks at the notes they receive, tweaks them a little, and passes them on. Nobody knows the whole answer alone, but together the team turns "raw photo" into "that's a cat". Training is the team practising on thousands of labelled photos and slowly agreeing on how much to trust each note. The weights are how much each person trusts the note in front of them.
How it learns
A fresh neural network starts with random weights, so its first guesses are nonsense. Training fixes that in a loop. The network sees an example, makes a prediction, and compares that prediction to the correct answer. The gap between the two is the error.
The network then nudges its weights a little in the direction that would have made the error smaller. This adjustment runs backwards through the layers, which is why the method is called backpropagation. Repeat this across millions of examples and the weights settle into values that produce good answers on data the network has never seen.
Two things matter for quality. The first is data: the network can only learn patterns that exist in its examples. The second is balance. Train too little and it stays vague; train too narrowly and it memorises the examples instead of learning the general pattern, which is called overfitting.
Where it is used
- Image and video recognising faces, reading scanned documents, spotting defects on a production line.
- Language translation, summarisation, and the large language models behind ChatGPT, Gemini, and Claude.
- Speech voice assistants and live transcription.
- Recommendations the suggestions you see on Netflix, Spotify, or an e-shop.
- Forecasting and detection predicting demand or flagging suspicious payments as fraud.
Common misconceptions
- A neural network does not think like a brain. The brain analogy explains the structure, not the experience. There is no understanding or intent inside, only weighted math repeated at scale.
- More layers is not automatically better. A bigger network needs more data and computing power, and it can overfit. The right size depends on the problem.
- It is not magic, and it is not neutral. A network learns whatever is in its training data, including the mistakes and biases. Garbage in, garbage out still applies.
- It can be confidently wrong. Like any statistical model, it produces a best guess, not a guaranteed truth. Verify anything that matters.
Related articles:
- What is an LLM? - The large language models behind ChatGPT and Claude are built on neural networks.
- Machine Learning vs Deep Learning - Where neural networks fit, and what makes a network "deep".
- What is AI? - The most searched terms around artificial intelligence, short and simple.
Want to stay one step ahead?
Don't miss our best insights. No spam, just practical analyses, invitations to exclusive events, and podcast summaries delivered straight to your inbox.