Learning and extraction of nonlinear mappings using an associative memory

Abstract

The problem of representation of nonlinear functions is considered using an associative memory structure, the associative memory network (AMN). AMN is a single layered neural network which uses input data to generate addresses of memory weights for learning and output of nonlinear functions. Within the framework of memory based learning of nonlinear mappings, several properties of AMN are analyzed through computer simulation and experiment. For example, the weight distribution in the course of learning of nonlinear functions is examined with respect to amplitude, time period, precision and offset of sampled input data. By doing so, generalization and specialization capability of AMN as well as robustness of learning to discretization level of input data are demonstrated.