Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!howland.reston.ans.net!math.ohio-state.edu!caen!msunews!harbinger.cc.monash.edu.au!yoyo.aarnet.edu.au!frodo.cc.flinders.edu.au!calvin!davidh
From: davidh@ist.flinders.edu.au (David Howard,Ph3751,Rm368)
Subject: Re: Gradient Techniques
Sender: @frodo.cc.flinders.edu.au
Message-ID: <1995Feb18.053940.34227@frodo.cc.flinders.edu.au>
Date: Sat, 18 Feb 1995 05:39:40 GMT
Reply-To: davidh@ist.flinders.edu.au
References: <3hmk3l$g7g@cantaloupe.srv.cs.cmu.edu>
Organization: Flinders University of S.A.
Lines: 37

>In general, there is no reliable relation between E' 
>and the curvature, so you can't do a good job of choosing
>the step size just from that.

How 'bout using finite differences in E' to approximate
E''?  Various routines described in "Numerical Recipes in C"
mention this.  Surely you can do a good job using this
method?  It's likely to be incredibly slow given that E' is
usually time consuming to compute, but training is done off
line anyway, and if it helps to get better weights ...

Comments?

Cheers,

David.

---
David Howard

Discipline of Mathematics
School of Information Science and Technology
The Flinders University of South Australia

PO Box 2100 
Adelaide  SA  5001
Australia

e-mail: davidh@maths.flinders.edu.au
phone:  (08) 201 3751


There we stood, two against a thousand ...

    ... and they were the toughest pair we ever fought!


