Path: utzoo!utgpu!news-server.csri.toronto.edu!mailrus!cs.utexas.edu!usc!snorkelwacker!ira.uka.de!fauern!tub!b21!fechner
From: fechner@b21.UUCP (Thomas Fechner)
Newsgroups: comp.ai.neural-nets
Subject: Neural Net Analysis Tools
Keywords: Analysis Tools, Backpropagation, Pattern Recognition
Message-ID: <406@b21.UUCP>
Date: 17 Aug 90 14:06:53 GMT
Organization: Daimler-Benz Research Institute Berlin, Germany
Lines: 21

I am working on neural nets for pattern recognition in signal processing 
applications. The first results with backpropagation training are (as
usually) very promising but there still remains one main problem:
After having trained the network, what are the discriminant features within 
the neural network ? Is there any tool which finds these automatically ? 
I know that Sejnowski analyzed his sonar classification network. But
he did it manually. For large networks there is no realistic way to do the 
trained-network analysis manually.

I would appreciate if somebody could suggest any ideas or related
references.

Thomas

--------------------------------------------------------------------------------
| Thomas Fechner 				Daimler Benz AG                |
| Parallel Processing Group 			Research Institute Berlin      |
| E-mail: uunet!mcsun!unido!b21!fechner 	Alt-Moabit 91b 		       |
| Phone: (..49-30) 39982-255 			D-1000 Berlin 21               |
| Telefax: (..49-30) 39982107			Federal Republic of Germany    |
-------------------------------------------------------------------------------