Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!rochester!udel!news.mathworks.com!news.alpha.net!uwm.edu!vixen.cso.uiuc.edu!howland.reston.ans.net!swrinde!pipex!uknet!bcc.ac.uk!pc6-033.ion.bpmf.ac.uk!W.ElDeredy
From: W.ElDeredy@ucl.ac.uk (Wael El Deredy)
Subject: A question about Skeletonization
Sender: news@ucl.ac.uk (Usenet News System)
Message-ID: <W.ElDeredy.30.2F44BCD3@ucl.ac.uk>
Date: Fri, 17 Feb 1995 15:12:19 GMT
Organization: Institue of Neurology
X-Newsreader: Trumpet for Windows [Version 1.0 Rev A]
Lines: 26


In Mozer and Smolensky's Skeletonization paper, NIPS 1 pp. 107-115:

They used  a linear Error function  E= Sum(abs(target-output))
instead of  E= Sum(target-output)^2
to provide a better estimate of relevance when the outputs are close to their 
target values (of one). The relevance estimate being   d(E)/d(alhpa)     where 
alpha is the attentional strength.

Now, I can't see that this change will avoid the drawback they were trying to 
rectify, namely the relevance tending to zero as the error decreases, since 
their relevance estimate will still contain the term  output (1 - output) (the 
derivative of the sigmoid function)

Does anyone know what exactly was the objective of using the linear error 
function ?

ThanX

Wael 
Wael El Deredy                            W.ElDeredy@ucl.ac.uk
Dept. of Neurological Surgery             Tel +44 (0)171 837 3611 x 4169
Institute of Neurology - Queen Square     Fax +44 (0)171 278 5069
London WC1N 3BG, UK


