Path: utzoo!attcan!uunet!tut.cis.ohio-state.edu!zaphod.mps.ohio-state.edu!samsung!cs.utexas.edu!sun-barr!newstop!sun!imagen!atari!portal!portal!cup.portal.com!mgj
From: mgj@cup.portal.com (Mark Gregory Jurik)
Newsgroups: comp.ai.neural-nets
Subject: Eliminate BackPropagation Blues
Message-ID: <27433@cup.portal.com>
Date: 1 Mar 90 10:29:35 GMT
Organization: The Portal System (TM)
Lines: 27

Pre-publication report on BACKPERCOLATION

While BackPropagation creates output error gradients for each cell, BackPercolation is a second pass that creates,
or assigns, true error to each cell.  With this information,
each cell has a much better "understanding" of what it needs
to do.  Experimental results show amazing learning behavior
during the very difficult 8-2-8 encoder problem, such as :

       1. relative insensitivity to initial condintions
       2. no apparent vulnerability to local minima
       3. exponential error decrease almost from the very start
       4. very stable with hidden layers ... the learning rate
                factor can be a full 1.0 !!

If you are interested, I will mail you a copy of the experimental
results via U.S. Post.  Your request *must* include a self-
addressed stamped envelope (standard business size).  I'll go
broke otherwise.  If you like what you see, I'll then send you
the algorithm details.

Why this approach and not via a conference?  Simply put, that
route will take too long.  I'm looking for co-researchers and/or
other experimenters NOW.  

Send your request to Mark Jurik
                     PO 2379
                     Aptos, CA  95001   USA