A cascaded artificial neural network architecture with novel robustness

Two 2-layer networks are first trained independently by delta rule and then cascaded. The middle layer can be viewed as a hidden layer and is trained to attain preassigned saturated outputs in response to the training set. Simulation results reveal that generalization ability of this cascaded network is far better than that of conventional back-propagation networks. Suggestions about the hidden coding in the cascaded network are presented. This network also learns considerably faster than BP networks. In large scale integrated neural network systems this network would enhance the overall performance