Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!bb3.andrew.cmu.edu!newsfeed.pitt.edu!news.duq.edu!newsgate.duke.edu!news.mathworks.com!howland.erols.net!EU.net!usenet2.news.uk.psi.net!uknet!usenet1.news.uk.psi.net!uknet!uknet!warwick!nott-cs!trent-doc!cgo
From: cgo@doc.ntu.ac.uk (Mr Chris Goodwin)
Subject: Can RMS go up?
X-Nntp-Posting-Host: bob
Message-ID: <Dywq42.Es0@doc.ntu.ac.uk>
Sender: news@doc.ntu.ac.uk
Organization: The Nottingham Trent University, DOC.
X-Newsreader: TIN [version 1.2 PL2]
Date: Mon, 7 Oct 1996 13:14:25 GMT
Lines: 20

Hello all,

	Please help solve the following puzzle:

	If you have a feedforward neural network which
is being trained with backpropagation, and you have
a zero momentum term, and you use on-line
training as opposed to batch mode, can the
RMS ever go up?

	A possible aid to thinking about this problem
is: If you are at the global minimum after an epoch, but
the error is not zero, where will the RMS and the
weight matrix go next?

Thanks for any contributions to our discussions,

Chris Goodwin
cgo@doc.ntu.ac.uk

