Neural network

A neural network lays out small units of computation (neurons) in several layers and connects them. Each connection multiplies the number it receives from the previous layer by a set strength, adds them up and passes the result to the next layer. After passing through many layers, a photo becomes the answer "cat," and a question becomes a reply.

At first the connection strengths are random, so the answers are nonsense. Compare the output with the correct answer, adjust the strengths a little by the amount of the error, and repeat this many times, and the answers get better. This is learning, and each connection strength is a parameter. Despite the name, it does not work exactly like a real brain.

It is like a game where people stand in a line and whisper a message along it. Each time a wrong message comes out at the end, you work out who passed it on badly and by how much, and fix it. In time, the message is passed on accurately.

Language models, image generation AI and self-driving all run on neural networks. Words in AI news like "layers," "weights" and "parameters" all come from here.