How do neural networks work?

A neural network is just multiply, add and squash, done millions of times with weights it learns from examples. An artificial neuron multiplies each input by a weight, adds a bias, and passes the total through an activation function such as a step, a sigmoid or ReLU.

Read how it works: How do neural networks work? · The history of neural networks · More boxes on Glassbox

Chapters in this interactive model

1. One artificial neuron: Multiply, add, squash: the tiny sum inside every neural network.
2. Layers: why one neuron is not enough: Each hidden neuron draws a line. Together, their lines bend.
3. Learning: rolling downhill: Measure the error, find which way is down, take a small step. Repeat thousands of times.
4. Seeing images: pixels in, guesses out: Draw a digit. A network trained here, in your browser, has a go.
5. Overfitting, and why data matters: A network can memorise instead of learn, and it can only be as fair as its data.
6. Where they are used, and their limits: Pattern finders everywhere: useful, powerful, and not the same as understanding.