Neural networks were loosely inspired by the brain, but they are really mathematical functions. Stack many layers and you get deep learning, the approach behind modern image recognition, speech recognition and large language models.
How it works
Each artificial neuron takes several numbers as input, multiplies each by a weight, adds them up with a bias, and passes the total through an activation function, which lets the network model curved, non-linear relationships. Neurons are arranged in layers: an input layer receives data such as pixel values, hidden layers transform it step by step, and an output layer gives the result, such as a probability that an image shows a cat. Training starts with random weights. The network makes predictions, a loss function measures the error, and an algorithm called backpropagation works out how much each weight contributed to that error. Gradient descent then nudges every weight slightly in the direction that reduces it. Repeated over many examples, the network learns useful features by itself, for instance edges, then shapes, then objects in images. Different architectures suit different data: convolutional networks for images, recurrent networks for older sequence work, and transformers for language and much else today. The perceptron, an early single-layer version, dates back to 1958.
Why it matters
Neural networks power most of the AI tools creators and businesses use, including transcription, background removal, image and video generation, voice cloning and chat assistants. Knowing how they learn explains their strengths and weaknesses: they are excellent at spotting patterns in large data, but need lots of examples, can absorb biases and are hard to inspect. That understanding helps you set realistic expectations and ask better questions of AI vendors.
Where you will see it
- Recognising faces or objects to tag photos automatically.
- Converting speech in videos and meetings into text.
- Removing or replacing backgrounds in images and video.
- Generating text, images, music and voices in generative AI tools.
Limits and common misconceptions
- Despite the name, neural networks do not think like brains; artificial neurons are greatly simplified maths.
- They often need large amounts of data and computing power to train well.
- Their internal workings are hard to interpret, so it can be unclear why they made a particular decision.
How it differs from related terms
- Neural network vs Deep learning A neural network is the model structure; deep learning is the practice of training neural networks with many layers.
- Neural network vs Transformer A transformer is one specific neural network architecture, built around attention, that powers today's large language models.
Frequently asked questions
How does a neural network learn?
It makes a prediction, measures how wrong it was, then adjusts its internal weights to be a little less wrong next time. This is done using backpropagation and gradient descent, repeated across thousands or millions of examples until the network's predictions become reliably accurate on data it has not seen.
What is the difference between a neural network and deep learning?
A neural network is the model: layers of connected artificial neurons. Deep learning refers to using neural networks with many hidden layers, often dozens or more, trained on large datasets. All deep learning uses neural networks, but a small network with one hidden layer is not usually called deep.
Are neural networks like the human brain?
Only loosely. The idea of connected units passing signals came from biology, but artificial neurons are simple mathematical functions, and networks learn through methods like backpropagation that brains are not known to use. Brains are far more energy-efficient and learn from far fewer examples.
Related terms
- Multimodal AI AI that can take in or produce more than one type of data, such as text, images, audio and video, and connect information across them.
- Natural language processing (NLP) The field of computer science and AI concerned with enabling computers to process, understand and generate human language, written or spoken.
- No-code Building apps, websites or automations through visual editors, templates and settings instead of writing programming code.
- Open-weight model An AI model whose trained weights are publicly released, so anyone can download, run and often fine-tune it on their own hardware.
- Parameters The numeric values inside an AI model, mainly weights and biases, that are adjusted during training and together store what it has learned.