What is a neural network? Layers of simple multiply-and-add units wired together: the backbone of modern AI

Layers of simple multiply-and-add units wired together: the backbone of modern AI.

A neural network is a way of computing loosely inspired by nerve cells in the brain. Each unit, or neuron, does something simple: it multiplies the numbers coming in by weights, adds them up, and passes a signal on if the total is high enough.

Thousands or millions of these units are wired together in layers. Training a network just means adjusting the weight on each connection. It isn't a copy of a real brain; it borrowed the name and the basic idea. Describing neurons as math goes back to a 1943 paper.

Think of it like this

Imagine your class passing a note down a line. Each person adds a score using their own rule and hands it on. Nobody does anything clever alone, but what comes out at the end of the line can be a pretty smart decision.

For example

A classic beginner exercise is a network that looks at a handwritten digit and says which number, 0 to 9, it is.

Common mix-ups

Mix-upA neural network is an exact copy of the human brain.
ActuallyEven the first artificial neuron model, in 1943, drew a nerve cell very simply as an on/off logic switch. Real brain cells fire electrical pulses over time, and their connections keep changing while they're in use.
Mix-upTraining adds new parts to the network.
ActuallyA model file is the same size before and after training. Only the values change, not how many numbers there are — yet a 1986 study showed that this alone lets the middle layers learn, on their own, to pick out the features the task needs.

Try it in class

For teachers at the start of a lesson — or for students working on their own.

Question: Each unit is simple — so why does a chain of them make complex decisions?
Activity: Play pass-the-note. Five people stand in a line; each changes the score by their own rule (say, "double it and add 1") and passes it on. Change one person's rule and see how the final score moves — then talk about how that's like training adjusting the weights. Open the 1943 record (E0001) where the idea began.