Shapes of Non-monotonous Activation Functions in Chaotic Neural Network Associative Memory Model and Its Evaluation

Abstract

The purpose of this paper is to investigate the performance of the associative memory model using Aihara's chaotic neural network with different activation functions. Sigmoid function, a monotonous function, was used in Aihara's original model. However, in the static associative memory, it is reported that the storage capacity of the network is improved when a non-monotonous function is used as the activation function. To improve the associative ability of chaotic neural network, kinds of non-monotonous functions have been proposed to serve as activation function. This paper investigates their difference as to retrieval ability, and proposes an advanced non-monotonous function. By computer simulation, we discuss about what kind of shape is good to improve the associative ability of chaotic neural network.