(Slices through) Weight Space

[pollack@cis.ohio-state.edu]

****** Do not forward to other b-boards or mailing lists.  thank you ****

This tech report, with plenty of pretty pictures, addresses the
relationship between initial and final points in weight space...

Jordan

---------------------------------------------------------------------------
		Back Propagation is Sensitive to Initial Conditions

				John F. Kolen
			      Jordan B. Pollack 

                            Report 90-JK-BPSIC
		Laboratory for Artificial Intelligence Research 
		  Computer and Information Science Department
			The Ohio State University 
			Columbus, Ohio  43210, USA 
			kolen-j at cis.ohio-state.edu
			pollack at cis.ohio-state.edu

				Abstract 

This paper explores the effect of initial weight selection on feed-
forward networks learning simple functions with the back-propagation
technique.  We first demonstrate, through the use of Monte Carlo
techniques, that the magnitude of the initial condition vector (in
weight space) is a very significant parameter in convergence time
variability.  In order to further understand this result, additional
deterministic experiments were performed.  The results of these
experiments demonstrate the extreme sensitivity of back propagation to
initial weight configuration.

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

This tech report is available by the usual method of anonymous FTP from
cheops.cis.ohio-state.edu in pub/neuroprose as

		kolen.bpsic.tr.ps.Z
		kolen.bpsic.fig1.ps.Z
		kolen.bpsic.fig2.ps.Z
		kolen.bpsic.fig3.ps.Z
		kolen.bpsic.fig4.ps.Z
		kolen.bpsic.fig5.ps.Z

Or, write for Report 90-JK-BPSIC to:

Technical Report Librarian
Laboratory for AI Research
Ohio State University
2036 Neil Ave.
Columbus, OH 43210