Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!news.mathworks.com!udel!gatech!swrinde!ihnp4.ucsd.edu!newshub.nosc.mil!news!news
From: "R. Scott Starsman" <r_starsman@nise-p.nosc.mil>
Subject: Re: Biasing
Message-ID: <1995Feb3.135845.26353@nosc.mil>
Sender: news@nosc.mil
Organization: NCCOSC RDT&E Division, San Diego, CA
References:  <2095@kechb.demon.co.uk>
Date: Fri, 3 Feb 1995 13:58:45 GMT
Lines: 26

Jon@kechb.demon.co.uk (Jon) wrote:
>
> Could someone explain to me why you need biasing
> in BP nets. I've tried with and without but I cant
> really say if it has made any difference with simple
> net configurations.
> 
> Thanks
> 
> Jon (Jon@Kechb.Demon.Co.UK)
> 

Pardon my previous blank reply.  Having a little trouble with my
news reader. 

Consider a simple 1-input, 1-output net (any number of hidden elements).

Assume that you wish to teach this net to perform a logical NOT 
(input a 1 and return a 0 and input a 0 and return a 1).

Without a bias neuron, this network can never return non-zero values 
when a zero is input.  The bias neuron(s) provide the network some input
to chew on even when all the inputs are zero.

Scott
r_starsman@nise-p.nosc.mil
