Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!bb3.andrew.cmu.edu!newsfeed.pitt.edu!portc02.blue.aol.com!howland.erols.net!cs.utexas.edu!newshost.convex.com!newsgate.duke.edu!interpath!news.interpath.net!news.interpath.net!sas!newshost.unx.sas.com!saswss
From: saswss@hotellng.unx.sas.com (Warren Sarle)
Subject: Re: Scaling Question
Originator: saswss@hotellng.unx.sas.com
Sender: news@unx.sas.com (Noter of Newsworthy Events)
Message-ID: <E2xM0n.2r1@unx.sas.com>
Date: Tue, 24 Dec 1996 18:57:59 GMT
X-Nntp-Posting-Host: hotellng.unx.sas.com
References: <32AA4E4C.7ABB@netreach.net> <32AC2994.68D1@ais.net>
Organization: SAS Institute Inc.
Lines: 32


In article <32AC2994.68D1@ais.net>, uthed@ais.net writes:
|> Joseph Bush wrote:
|> > 
|> > I have been training and evaluating NNs using scaling of 0 to 1 for my
|> > inputs.  All my inputs and outputs are variables data and I am using a 3
|> > layer backprop network.  What differences would I notice if I used
|> > scaling -1 to 1.  What guidance can people give as to when to use what
|> > scaling technique?
|> > 
|> > Joe
|> 
|> Probably not much. If your inputs contain no theoretical negative
|> values, scale from zero to one. If they do, scale from minus one to
|> positive one.

I don't see any reason why the range of the original variable should
affect the choice of scaling/standardization for an input.

For a typical standard backprop net, scaling inputs to [-1,1] will
give you a better chance of finding a global optimum and faster
learning than scaling to [0,1]. I have elaborated the FAQ somewhat
in this regard--see "Should I normalize/standardize/rescale the data?"
in the Neural Network FAQ, part 2 of 7: Learning, at 
ftp://ftp.sas.com/pub/neural/FAQ2.html
-- 

Warren S. Sarle       SAS Institute Inc.   The opinions expressed here
saswss@unx.sas.com    SAS Campus Drive     are mine and not necessarily
(919) 677-8000        Cary, NC 27513, USA  those of SAS Institute.
 *** Do not send me unsolicited commercial or political email! ***

