Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!bb3.andrew.cmu.edu!newsfeed.pitt.edu!newsflash.concordia.ca!news.nstn.ca!ott.istar!istar.net!van.istar!west.istar!n1van.istar!van-bc!nntp.portal.ca!newsfeed.direct.ca!hunter.premier.net!op.net!news.mathworks.com!newsfeed.internetmci.com!usenet.eel.ufl.edu!warwick!lboro.ac.uk!usenet
From: Dave the Troll <eldb3@lboro.ac.uk>
Subject: Re: Assessing the importance of network inputs...
Sender: usenet@lboro.ac.uk (Usenet-News)
Message-ID: <31EE03CD.36D@lboro.ac.uk>
Date: Thu, 18 Jul 1996 09:28:45 GMT
X-Nntp-Posting-Host: elpc4726.lut.ac.uk
Content-Transfer-Encoding: 7bit
Content-Type: text/plain; charset=us-ascii
References: <31E4D647.794B@uea.ac.uk> <4siori$hpi$1@mhadg.production.compuserve.com>
Mime-Version: 1.0
X-Mailer: Mozilla 2.01 (Win95; I)
Organization: Loughborough University, UK.
Lines: 29

Somebody wrote:
> I wish to assess the relative importance of my network input
> variables in order to get a better understanding of the
> relationship between them and the network output.  Is there any
> particular way that this is carried out - any papers which have
> used looked at this ?

In the paper cited below we discuss a method by which a sensor is 
modelled using an MLP.  The sensor (MLP) is initially assumed to have 
many measurands (inputs).  The sensitivity of the input/output 
relationship is computed by calculating the second deviative of the one 
wrt the other.  The least sensitive is remove and the network retrained. 
 The process is repeated until a minimal sensor has been designed.

This technique is similar to some neuron pruning algorithms, with the 
exception that it is the inpout layer that is being pruned.

I hope that this is of some use,
  Dave Barnett
  Optical Engineering Group
  Dept. of Elect and Elec Eng
  Loughborough University

http://info.lut.ac.uk/departments/el/research/optics/

Paper -
Naimimohasses R, Barnett DM, Green DA, Smith PR, "Sensor optimization 
using neural network sensitivity measures." Meas. Sci. Technol. 6 (1995) 
 pp 1291-1300
