Associative memories are composed of a collection of interconnected elements having data storage capabilites. The elements are accessed in parallel by the content of a data probe vector rather than by a specific address. Recently, models referred to as the Hopfield AMN have received much attention. These models are simple to analyze and exhibit some of the behavioral features of neural networks. To incorporate these features, some new models will be explored which retain enough of the simple structure of the Hopfield AMN that makes them amenable to analysis. The objective is to obtain a better understanding of the properties and behavior of various types of AMN. The principal investigator is analyzing the dynamic and information storage capabilities of several new models. Many of the new models are subject to some physical constraints and interconnection costs, and it is appropriate to formulate various network design and analysis problems as optimization problems. The research will have important implications for the use of AMN in computer architectures of the future, matched to challenging problems in signal processing, communications, control, and artificial intelligence. Understanding the relationship between parameters of various AMN and their information storage capacities is a step towards determining the most useful aspects of this form of parallel distributed computing.
