Attractor Analysis in Associative Neural Networks and its Application to Facial Image Analysis

Autoassociative neural networks can be used for nonlinear processing and normalization of data. Because, firstly, they are able to learn and simulate complex nonlinear communications, and secondly, the communications can be learned through analyzing and distributing information on the neurons and weights and then combining the results of their processing. In this way, they actually make an interpolation between the input data and their communications. But these neural networks cannot model attractor dynamics that is obviously used in brain function. In this paper, the output of autoassociative neural network is connected to its input, and through recursive connections the ability of attractor behavior in these models is provided. This study showed that a recursive neuron with a logistic function forms two attractors, in its training point and its symmetry, but for it with a sigmoid nonlinear function can be formed an attractor in a certain range. In the experiments on face images, it was shown that the absorbance of the images to their attractors was improved from 52.67% to 87.27% by increasing the number of layers and the supervised layer-by-layer pre-training in order to adjust the attractors.