Two connected steps

In 1986, David Rumelhart, Geoffrey Hinton, and Ronald Williams published an influential demonstration of this method. Their networks learned useful internal features because the final error could guide weights in earlier layers.

The paper did not make deep networks easy to train. It made the route from a mistake to an earlier weight clear enough to use.

You can let the smaller calculation details go for now. Keep two verbs.

Keep two verbs:

Backpropagation calculates.

Gradient descent adjusts.

Then the network tries again.

Those two steps repeat across many examples. The network gets an answer wrong, traces the error backward, adjusts, and tries again.