What is Observable in a Class of Neurodynamics?

摘要

Nearly all models in neural networks start from the assumption that the input-output characteristic is a sigmoidal function. We study a class of dynamical systems of neural network models with saturated sigmoidal functions as their input-output characteristics. A complete spectrum on the parameter space for all possible outcomes of the dynamics is obtained. Under a stated condition we show that the possible outcomes of the dynamics are all saturated or all-but-one saturated fixed point at tractors. An exact parameter region is given for all saturated attractors and all-but-one saturated fixed points.