Technical Report Available (was Fractal Representations)

[John Merrill]

The following technical report is available from the Indiana
University Department of Computer Science as TR #249.    

Title: Fractally Configured Neural Networks
Authors: Merrill, John W. L. adn Port, Robert F.

			       ABSTRACT

The effects of network structure on learning are investigated.  We
argue that there are problems for which specially tailored network
structures are essential in order to achieve a desired result.  We
present a method to derive such network structures, and present the
results of applying this algorithm to the problem of generalization in
abstract neural networks.  In order to derive these networks, it is
essential that the system employ a flexible, yet efficient,
representation of edge structure.  The algorithm discussed here uses
deterministic chaos to generate a fractal partition of the edge space,
and uses that fractal partition to produce an edge structure.  We
discuss the results of applying this algorithm to a simple
classification problem, and we compare the performance of the
resulting network to the performance of standard feed-forward
networks.  Our results show that the specially constructed networks
are better able to generalize than completely connected networks with
the same number of nodes.

Electronic requests should be sent to merrill at bucasb.bu.edu on ARPA or
to port at iuvax on UUCP.  Physical requests should be sent to:

	Nancy Garrett
	Department of Computer Science
	101 Lindley Hall
	Indiana University
	Bloomington, Indiana  47405

------------------------------------------------------------------------
John Merrill			|	ARPA:	merrill at bucasb.bu.edu
Center for Adaptive Systems	|	
111 Cummington Street		|	
Boston, Mass. 02215		|	Phone:	(617) 353-5765