A network can see that it predicted the wrong answer and still face a harder question. Which of the many connections inside it should change? In a 1986 Nature paper, David Rumelhart, Geoffrey Hinton and Ronald Williams showed how a learning procedure called backpropagation could answer that question for a network with hidden layers.
A hidden layer sits between the inputs and the final answer. It transforms the information the next layer receives. Backpropagation starts with a measure of the output error and works backward through the network to calculate how small changes to the connections would affect that error.
The middle of the network could learn something useful
Adjusting the internal connections lets a network learn intermediate patterns that help with its task. The 1986 paper illustrated that behavior in small examples. The changes emerged from the objective used to train the network, without requiring people to write a rule for each pattern the hidden units should learn.
That helps explain why the paper became influential. Training could shape what happened between the input and the answer. The examples show the method at work long before today's enormous networks and datasets.