Backpropagation is a supervised learning for multi-level neural networks. For each training example, it takes the difference between the expected and actual output at the final layer and then uses the differential of the sigmoid function at each node to work out error values at earlier layers and hence update the weights on links between nodes.
Used in Chap. 6: pages 84, 87; Chap. 7: page 105; Chap. 8: pages 111, 113; Chap. 9: pages 132, 133, 137, 139; Chap. 12: page 196; Chap. 20: page 344; Chap. 21: page 357
Also known as backprop