A weighting approach for autoassociative memories to improve accuracy in memorization

An autoassociative memory can store multiple information in a neural network, and if some distorted information is presented, the memory can retrieve the most likely information from the network. However, in mathematical models of the autoassociative memory, it is a significant problem that some given information may not be stored correctly in a recurrent artificial neural network (ANN). In this paper, in order to investigate the cause of errors with memorization rules in such a mathematical model, we understand the structure of the energy function for the ANN as a sum of elemental quadratic functions. Then, in order to improve the accuracy in memorization, we propose a weighting approach for the memorization rules so that the structure of the energy function can be altered in a desirable manner. The weights can be determined by solving a theoretically-derived linear program to guarantee perfect memorization of all the given information. Numerical examples demonstrate the effectiveness of the weighting approach.