Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!rochester!cornellcs!newsstand.cit.cornell.edu!ub!csn!news-1.csn.net!magnus.acs.ohio-state.edu!math.ohio-state.edu!uwm.edu!newsfeed.internetmci.com!in1.uu.net!bcstec!nntp
From: Mike Fouche <mikef@hsvaic.hv.boeing.com>
Subject: Re: Help: Improvement of Generalization of Feedforward ANN
X-Nntp-Posting-Host: e677744.hv.boeing.com
Content-Type: text/plain; charset=us-ascii
Message-ID: <DoM6v0.5AA@bcstec.ca.boeing.com>
Sender: nntp@bcstec.ca.boeing.com (NNTP News Access)
Content-Transfer-Encoding: 7bit
Organization: Missiles & Space Division
References: <4ih47a$gn5@wmwap1.math.uni-wuppertal.de> <DoHGLB.IDL@bcstec.ca.boeing.com> <4imi01$c4l@wmwap1.math.uni-wuppertal.de>
Mime-Version: 1.0
Date: Thu, 21 Mar 1996 11:06:34 GMT
X-Mailer: Mozilla 1.2 (Windows; U; 16bit)
Lines: 22

You indicated 10 real values as input but later say that 1000 are needed.  I got
lost somewhere and don't understand where the 1000 input values comes from.  Unless
you mean that you convert a number, say 125, to an integer so that one input becomes
7 inputs (binary of 125 (one input) = 1 1 1 1 1 0 1 (seven inputs)).  Then 10 inputs
x 7 = 70 inputs to the net.  Still less than 1000.  Do you mean 1000 overall data
inputs?  80 training sets x 70 inputs = 5600 data points => is this what your 
saying?

The small data (1e-25 should definitely be made MUCH larger - else you risk losing
accuracy due to machine precision - I believe ...).  If your data varies between
say 1e-25 and 1e-10 or something like that then consider using log10 scaling and 
scale it in such a way that you "set the widest window" around the data of most
importance.  We typically use for some applications a log10 or ln scaling then do
a linear mapping.

If your input to the net is only 70 inputs, then I don't think you need FFT, but if
it really is 1000 then it would be worthwhile.  There are several ways to do an FFT,
such as 1-D or 2-D transform.  "Numerical Recipes in C" has a good section on FFTs
and easily implementable C code (very portable).

Mike Fouche

