Discrete N-Layer Heteroassociative Memory

Rodica Waivio (joint work with Bhaskar DasGupta)
Department of Computer Science, University of Illinois at Chicago

Abstract 

In this paper a new generalized N-Layer HeteroAssociative
Memory Model is proposed for discrete case. The recurrent model is
an improvement and an extension of discrete BAM and Hopfield
models to N-layers. Based on pattern decomposition the standard
form is determined. After a competitive initialization between all
layers the multilayer network converges in one step to fixed
points which are standard memories or are very close to these. By
increasing the domain of attraction the architecture is much
powerful than the previous models. An analysis of the optimal
number of layers as well as their dimension is realized based on
existence of subspaces in the pattern. The stability of the new
model is proved using Lyapunov theory under specified bounding
hypotheses. The theory proposed and some experiments show that
this model improves the recall process much better than other
existent models.