Analysis of Limit-Cycles on Neural Networks with Asymmetrical Cyclic Connections Using Approximately Activation Functions

摘要

We analyze the limit-cycles generated on the neural network with asymmetrical cyclic connections in continuous-time model. We analytically obtain the relations between the period of these limit-cycles and various parameters in the network using two kinds of approximately activation function, linear and high-gain limit. From consideration of these results, nonlinear characteristics appears to produce the region of attraction to the solution that has a main single-frequency component in various solution space that consists of the superposed solutions of the linear approximation.