A feedforward neural network behaving as a hierarchical system of automata

Abstract

A new design principle of a feedforward neural network is presented. The built net behaves as a hierarchical system of automata where each action of an automaton at a certain level triggers an automaton at the level directly below it. The activation of hidden nodes is governed by the excitation degree of the input layer. The learning algorithm, based on the gradient search method, updates only the selected hidden node parameters. The studied network reduces the computational load in both training and operating steps. Simulation results show the effectiveness of the proposed network compared with the conventional feedforward neural network.