Generalization ability of artificial neural network using Fahlman and Lebiere´s learning algorithm

The authors investigate the generalization ability of the network generated by S.E. Fahlman and C. Lebiere´s (FL) (1990) learning algorithm which has a distinctive feature to build any network topology. They trained the same three-layer network by the FL algorithm and the backpropagation (BP) algorithm, and a comparison of recognition abilities shows that the FL network performs much better than the BP network. The FL network performs better because, in this network, hidden units use only saturated values and thus the hidden layer acts as a filter for noise. A two-layer network performs excellently if the training set is trainable by the two-layer network and if a pattern is recognized by detecting the maximum valued output. Since the FL algorithm begins with a minimal two-layer network which performs best under the condition stated, a designer can construct either a two-layer or a multilayer network according to which one best fits a particular application. Thus, it can be concluded that in all these respects the FL algorithm is preferable to the BP algorithm