Training hard-limiting neurons using back-propagation algorithm by updating steepness factors
Xiangui Yu, N.K. Loh, W.C. Miller · 1994
This paper presents one kind of modified backpropagation algorithm for training the multilayer feedforward neural networks with hard-limiting neurons. Adaptive steepness factors in the analog sigmoidal neuron activation functions are updated in the training process. With the decrease of the sum-square error, these steepness factors are varied from a small positive value to infinite. It makes the sigmoidal neuron transferred to hard-limiting one after the training process complete. Thus, a multilayer feedforward neural network can be trained with the resultant architecture is only composed of hard-limiting neurons. The learning algorithm is similar to the conventional backpropagation algorithm, only the derivatives of the hidden neural activation functions are modified according to the proposed idea. Extensive numerical simulations are presented to show the feasibility of the proposed algorithm. In addition, the numerical properties of the proposed algorithm are also discussed in detail. Comparisons of the proposed algorithm with algorithms are given, and some useful conclusions are drawn.>