AI Trend Research

← AI glossary

AI glossary · Beginner · Language models

Artificial neural network

An artificial neural network is a computing structure that connects small units that take numbers and compute (neurons) in multiple layers, and learns rules by adjusting the strength of those connections on its own as it sees data. It is the basic backbone of today's deep learning and LLMs.

About this term
Level
Beginner · Core ideas for first-time readers
Category
Basics
Added
2026-10-08
In plain words It is similar to a company where many employees sit in tiers, each slightly processing the memo received from the person in front and passing it to the person behind. Each employee learns the job by adjusting, through experience, how much to trust whom. However, it does not actually think like brain cells; it is mathematical computation that only borrows its name from the brain.

How it works

A neural network consists of an input layer, one or more hidden layers, and an output layer. Each neuron multiplies the numbers received from the previous layer by their own weights (connection strengths), adds them up, passes the result through a rule called an activation function, and sends it to the next layer. At first the weights are set randomly, so the network gives wrong answers, but it learns by repeatedly adjusting the weights little by little in the direction that reduces the gap from the correct answer. These weights are what are commonly called "parameters."

A neural network built from layersInput layer → Hidden layer 1(×weight); Hidden layer 1 → Hidden layer 2(×weight); Hidden layer 2 → Output layer(×weight)Input layerPhoto pixels, words,etc.Hidden layer 1Weighted sum +activation functionHidden layer 2More complex featuresOutput layere.g. probability of acatWeights (parameters)Adjusted little by little during training×weight×weight×weight
A neural network built from layers Numbers enter at the input layer on the left, are processed as they pass through the hidden layers in turn, and an answer comes out of the output layer on the right. Each arrow between layers has a weight that is adjusted during training.

Why it matters

Without people writing out rules one by one, it can learn complex patterns, such as objects in photos, speech sounds and the meaning of sentences, from data alone. The field that deals with neural networks with many stacked layers is deep learning, and CNNs for image recognition and RNNs and transformers for sentence processing are all forms of neural networks. Most generative AI in use today is built on this structure.

Good to know

Neural networks are sometimes called "black boxes" because it is hard for people to see why they gave a particular answer, and the research that tries to open them up is interpretability. Training requires a lot of data and computing resources (such as GPUs). If a network memorizes its training data too closely, overfitting can occur, where performance drops on data it has not seen before.

Example

When a smartphone photo app automatically sorts pictures into 'food', 'landscape' and 'people', or a voice assistant turns speech into text, an artificial neural network usually does most of the work. LLMs such as ChatGPT and Claude are also huge neural networks built on an architecture called the transformer.

Related terms

Perceptron Backpropagation Deep learning Activation function Parameter Transformer