paper available on dynamic pattern formation

[Lina Massone]

The following paper is now available as a technical report of the Neural
Information Processing Laboratory of Northwestern University. The paper
can be retrieved through the web by connecting to:
                       http://www.eecs.nwu.edu/pub/nipl
or by anonymous ftp to:
                       eecs.nwu.edu
                       cd pub/nipl

 
	The role of initial conditions in dynamic pattern formation 

		Lina L.E. Massone and Tony Khoshaba

	    Tech. Rep. Neural Information Processing Laboratory
			Northwestern University
			*Submitted to NIPS 95*

In this paper we present the results of an empirical study of the 
properties of recurrent backpropagation (Pineda, 1988) with special
emphasis on the characteristics of the resulting weight distributions
and their dependency on the initial conditions and on the classes of
dynamic tasks that the network learns. The results of this study indicate
that the weights of the trained network exhibit properties that are
dictated by both the desired equilibria and the initial values of the weights,
but not by the initial state of the network. In particular, we were able
to quantify the dependency of the final weights on the initial weights
in terms of monotonic, practically linear relationships between the standard
deviations of the two distributions. We discuss the implications of these 
results for dynamical systems in general and for the study of the brain
functions.