A mistake changes the weights

The Perceptron looks at a square and guesses: "circle."

A person supplies the correct answer: "square."

The correction procedure compares the guess with the correct answer. It changes some weights connected to the active parts of the image.

Then the square passes through again.

The image is the same. The weights are different. This time, the Perceptron answers: "square."

One correction is not enough to recognise every square. The process must repeat across many labelled examples.

Repeating this loop across many labelled examples is called training.