A Framework for Analysis of Granular Neural Networks

Julian Skirzynski

Abstract: Granular neural networks are neural networks which operate at the level of information granules. Granules, in turn, can be seen as collections of objects that exhibit similar structure or possess similar functionality. In this work we try to provide a comprehensive look at the problem of how granular, feed-forward neural networks conduct their computations, i.e. what is the interpretation for the connections and the neurons of such networks. The paper orbits around the assumption that the networks come from the superposition of their certain subnetworks which emulate membership functions for the granules. The superposition represents an aggregation of a certain number of granules into another one. This interpretation comes from a general granular tree-model that is constructed prior to the network and which describes a particular problem in a semantically understandable form.