Bump formations in attractor neural network and their application in image reconstruction.

In this paper we analyze the bump formations in binary attractor neural networks with distance dependent connectivity. We show that by introducing a two stage learning procedure an increase of the critical storage capacity of the network is observed. The procedure has been tested on a network with N = 64K neurons by using a selection of 3700 natural images. Our analysis shows that the bumps can be regarded as intrinsic characteristics of the image and the topology of the network and they can be used to improve the performance of the network by increasing its capacity. © 2007 American Institute of Physics [ABSTRACT FROM AUTHOR]