A new neuron model “cone” with fast convergence rate and its application to pattern recognition
Minoru Fukumi, Sigeru Omatu, Masaru Teranishi · Systems and Computers in Japan · 1991
Abstract This paper proposes a new neuron model “analog coupled neuron (a‐CONE)” with two sigmoid functions which is applied to parity and pattern recognition problems to show its ability. Also, a digital‐CONE (d‐CONE) capable of providing a digital output is proposed and it is trained by using a back propagation (BP)‐type algorithm. Moreover, it is shown that its speed of convergence in learning was faster than that of the standard BP. The a‐CONE presented here has two kinds of sigmoid functions f1 and f2 mapping an input into an output. The function f1 provides an output response of the neuron while the derivative in learning is given by both f1′ and f2′, where f1′ refers to the derivative of f1 with respect to weighted sum of input signals. This allows a fast convergence rate to be obtained in learning. From simulation results, it is shown that the a‐CONE has a rapid convergence rate in learning compared with the conventional BP and the learning algorithm of d‐CONE networks, and produces a noise‐tolerant mapping in the pattern recognition.