Report available in Neuroprose Archives

[steensj@daimi.aau.dk]

     ******  Please do not forward to other lists.   Thank you  *******


The following report has been placed in the neuroprose archives at Ohio State.
Ftp instructions follow the abstract.


       -------------------------------------------------------------


	         A Conceptual Approach to Generalization in
			   Dynamic Neural Networks

			       Steen Sjogaard
			 Computer Science Department
			      Aarhus University
			      DK-8000 Aarhus C.
				   Denmark
			    steensj at daimi.aau.dk


				  ABSTRACT

Inspired by the famous paper "Generalization as Search" by Tom Mitchell from
1982, a conceptual approach to generalization in artificial neural networks 
is proposed.  The two most important ideas are (1) to consider the problem of 
forming a general description of a class of objects as a search problem, and 
(2) to divide the search space into a static and a dynamic part.  These ideas 
are beneficial as they emphasize the evolution or process that a learner must 
undergo in order to discover a valid generalization.  We find that this 
approach and the adapted conceptual framework provide a more varied and 
intuitively appealing view on generalization.  Furthermore, a new cascade-
correlation learning algorithm which is very similar to Fahlman and Lebiere's 
Cascade-Correlation Learning Architecture from 1990, is proposed.  The 
capabilities of these two learning algorithms are discussed, and a direct 
comparison in terms of the conceptual framework is performed.  Finally, the 
two algorithms are analyzed empirically, and it is demonstrated how the 
obtained results can be explained and discussed in terms of the conceptual 
framework.  The empirical analyses are based on two experiments: The first 
experiment concerns the scaling behavior of the two network types, while the 
other experiment concerns a closer analysis of the representation that the two
network types utilize for found generalizations.  Both experiments show that 
the networks generated by the new algorithm perform better than the networks 
generated by the Cascade-Correlation Learning Architecture on the relatively 
simple geometric classification problem considered.


       -------------------------------------------------------------

To retrieve by anonymous ftp:

	unix> ftp archive.cis.ohio-state.edu
	Name: anonymous
	Password: neuron
	ftp> cd pub/neuroprose
	ftp> binary
	ftp> get sjogaard.concept.ps.Z
	ftp> quit
	unix> uncompress sjogaard.concept.ps.Z
	unix> lpr -P<printer name> sjogaard.concept.ps

/Steen