How does a neural network work?

Inspired by the brain

An artificial neural network is loosely inspired by the workings of biological neurons in the brain. Just like real neurons, artificial neurons are connected to each other — they receive signals, process them, and pass them on to the next layer.

But the similarity largely ends there. Artificial neural networks are mathematical models that work with numbers and functions, not electrochemical signals.

The building blocks: layers and neurons

A neural network consists of layers:

• Input layer — receives the raw data (e.g. pixels of an image or words in a sentence)
• Hidden layers — process the data in multiple steps; this is where the 'learning' happens
• Output layer — produces the final result (e.g. a category, a number, or a word)

Each neuron in a layer is connected to neurons in the next layer via weights: numbers that determine how strong a connection is. Those weights are adjusted during the learning process.

Illustration created with Canva AI

How does a neural network learn?

The learning process is called training and works as follows:

1. The network is presented with an example (e.g. a photo of a cat)
2. It makes a prediction (e.g. "dog")
3. The error is calculated via a loss function
4. Via backpropagation, all weights are slightly adjusted to reduce the error
5. This process repeats millions of times until the error is minimal

Activation functions

Each neuron applies an activation function to its input — a mathematical transformation that determines whether and how strongly a neuron 'fires'. Well-known activation functions are ReLU, sigmoid, and tanh. Without activation functions, a deep network would behave like a single linear function — too limited for complex tasks.

Deep learning

Deep learning simply means that the network has many hidden layers — sometimes hundreds. The deeper the network, the more complex patterns it can learn. This makes it possible to recognize faces, understand language, and play Go at the level of world champions.

Applications

• Image recognition (faces, medical scans, autonomous vehicles)
• Speech recognition (Siri, Google Assistant)
• Machine translation (DeepL, Google Translate)
• Recommendation systems (Netflix, Spotify)
• Large language models (GPT-4, Claude, Gemini)