When training a multi-layer neural network, the desired responses of the output layer neurons are given for each input training pattern. The desired responses for the neurons in the other layers are not known. The backpropagation algorithm generates desired responses for all the neurons in the “hidden layers,” the layers that feed signals to the output layer. Known errors at the output layer are backpropagated throughout the network to provide error signals for all the neurons of the network, thus enabling adaptation and learning for the entire network.