AI Search 101
What is a neural network?
A neural network is a system of simple computing units arranged in layers, each connected to the next by weighted links. Data passes through, each unit combines what it receives and passes on a result, and learning happens by adjusting the weights until the output is right.
The basics
| Also called | Artificial neural network, ANN |
| Loosely inspired by | Neurons in a biological brain, at a considerable distance |
| Built from | Layers of units connected by weights |
| Trained by | Backpropagation and gradient descent |
| Many layers means | Deep learning |
| Part of | Machine learning |
How does a neural network work?
Start with a single unit. It receives several numbers, multiplies each by a weight, adds them up, and passes the total through a function that decides what to output. That is the whole of it. One unit computes something trivial.
The capability comes from arrangement. Units are grouped into layers: an input layer that receives the data, one or more hidden layers, and an output layer that produces the answer. Every unit in one layer feeds units in the next, each connection carrying its own weight.
A network's behaviour is entirely contained in those weights. The same architecture with different weights recognises handwriting or predicts demand. Training means finding the weights that make the output correct, and nothing else about the network changes.
How does a neural network learn?
By being wrong repeatedly and correcting in small steps.
An example goes in and the network produces an output. That output is compared with the correct answer and the difference is measured as a loss. The question then is how much each weight contributed to the error, and the answer is computed by working backwards from the output to the input, layer by layer. This is backpropagation.
Each weight is then nudged slightly in the direction that would have reduced the error. This is gradient descent, and the size of the nudge is the learning rate. Repeat across millions of examples and the weights settle into a configuration that handles the task.
Nothing here resembles understanding. It is an enormous, patient optimisation, and what emerges is a set of numbers that happens to transform inputs into useful outputs.
Is a neural network like a brain?
Only in the loosest sense, and the analogy causes more confusion than it resolves.
Biological neurons communicate with timed electrical spikes, exist in enormous variety, form and dissolve connections physically, and operate in a body with chemistry and continuous sensory input. An artificial unit is a weighted sum followed by a simple function. The resemblance is a historical inspiration, not a description.
The name is worth treating with care in business conversations, because it invites people to attribute understanding, intention or awareness to a system that has none.
Why did neural networks take so long to work?
The mathematics was largely in place by the 1980s. What was missing was everything else.
Data. These systems need enormous quantities of examples, and the digitised world that could supply them did not exist until the 2000s.
Computation. Training requires vast numbers of simple parallel calculations, which is precisely what graphics processors do. Repurposing them in the late 2000s changed what was affordable by orders of magnitude.
Technique. A collection of practical advances made deeper networks trainable rather than unstable.
All three arrived together, which is why the field appeared to leap forward after decades of modest progress.
How neural networks are used in marketing
Production. Every generative tool a marketing team uses is a neural network underneath, whether it produces text, images or audio.
Analysis. Classifying open-text feedback, scoring propensity and churn, recognising objects in images, and forecasting.
Distribution. The ranking and recommendation systems inside advertising and social platforms are neural networks deciding which message reaches whom.
How this impacts your business
A language model does not look a company up in a register. It composes a description from everything it read during training and whatever it retrieves at the moment of the question. Trade press, directories, forums, competitor comparison pages, and somewhere in that mix the company's own website too.
So the description a buyer sees is assembled, not published. No company controls it, but every company supplies a share of the material it is built from.
The traffic impact is still small, roughly one percent of sessions for most companies. The influence is not small, because the reader is often a buyer forming a first impression of what a company does and who it serves, before any page gets a click.
Checking takes ten seconds. Type the company name into ChatGPT and read the answer.
Frequently asked questions
What are weights and parameters?
The numbers on the connections between units. Parameter counts quoted for large models, in the billions, are counts of these values. They are what training produces and what a trained model consists of.
What is a hidden layer?
Any layer between the input and the output. It is called hidden because its values are internal working state rather than something the user supplies or reads.
Can anyone tell what a neural network learned?
Only partially. Individual weights mean nothing in isolation and the representation is distributed across the whole network. Interpretability research makes progress but full explanation is not available for large models.
What is a transformer?
A neural network architecture, introduced in 2017, that lets every part of an input attend to every other part. It is the basis of current large language models.
Do neural networks need a graphics processor?
For training at scale, effectively yes. Small networks train on ordinary hardware, and many trained models run adequately without one.
Related definitions
- What is deep learning?
- What is machine learning?
- What is a large language model?
- What is artificial intelligence?
- AI Search 101
Written by Lari Numminen, Generate More. Also available in Finnish. Updated 31 August 2026.