A Subsethod Interval Associative Memory with Competitive Learning

Abstract

Morphological perceptrons with competitive learning (MP/CLs) are constructive artificial neural network models having a modular architecture. Not only the weights but also the architecture of the MP/CL is automatically generated by the MP/CL training algorithm. The resulting architecture is determined by hyperboxes, i.e., closed intervals, contained in Fn, where F is a totally ordered group. The group operations and the total ordering are used to perform a competition among the outputs of each module. In this paper, we present an interval subsethood associative memory whose hidden nodes compute degrees of subsethood of the input pattern in each of the closed intervals generated by the MP/CL training algorithm. We show that the resulting interval subsethood associative memory with competitive learning (S-IAM/CL) can be viewed as a ɵ-fuzzy associative memory (ɵ-FAM) model. We compare the performance of S-IAM/CLs for a particular choice of interval subsethood measure with the ones of the original MP/CL model and several competitive models in a number of classification problems.