Tri-State Neural Network and Analysis of its Performance

Abstract

ABSTRACTThe tri-state neuron is introduced to simulate the tri-state property of neurons in the nervous system. The double sigmoid function is utilized as the activation function of the tri-state neural network. The convergence rate and the generalization ability of the tri-state network are compared with those of the traditional bi-state network with single sigmoid function. It is shown analytically and computationally that (1) the tri-state network has a faster convergence rate than the bi-state network and that (2) the generalization ability of these two networks are comparable.