Bidirectional Associative Memory and Learning of Nonlinearly Separable Tasks

Abstract

Research has shown that human beings are able to perform various associations, whether linearly separable or nonlinearly separable, with little effort. While bidirectional associative memory (BAM) models show great promise in modeling various types of associations the humans perform, they still have difficulties with solving various types of nonlinearly separable problems. The present study introduces a modification of the architecture of a given type of BAM by adding an unsupervised pathway to the original BAM structure. Results showed that the modification allows the network to perform nonlinearly separable associations such as the n-bit parity task and the double-moon problem. The network is able to associate more difficult types of problems while keeping the same learning and transmission function. This study could lead to enhanced cognitive models capable of modeling a wider set of associations.


Back to Table of Contents