Path: utzoo!utgpu!news-server.csri.toronto.edu!bonnie.concordia.ca!uunet!mcsun!ukc!cf-cm!dkrm
From: dkrm@computing-maths.cardiff.ac.uk (Daniel K R McCormack)
Newsgroups: comp.ai.neural-nets
Subject: help with pb alg and zero weights
Message-ID: <1991Feb7.113526.18447@computing-maths.cardiff.ac.uk>
Date: 7 Feb 91 11:35:26 GMT
Sender: dkrm@computing-maths.cardiff.ac.uk (Daniel K R McCormack)
Reply-To: dkrm@computing-maths.cardiff.ac.uk (Daniel K R McCormack)
Organization: University of Wales College of Cardiff, Cardiff, WALES, UK.
Lines: 8

Hi , 
	This problem is most probably blantently obvoius to most neural netters except me . I have read somewhere that a backpropogation alg requires random weight initialisation , and that an all zero weight matricies will result in the network not learning. I have tried this out with a few simulation packages and they seem to learn with zeroed weights . 
If this is the case , how do you represent 'dead' connections ( i.e. connection that in theory are not present ) in the network . I've even looked at the equations for bp learning and I cannot see that a weight is not altered if the weight = zero . 
I really hope someone can put me out of my misery soon
Thanks in Advance 
	Daniel

email dkrm@cm.cf.ac.uk