Message-ID:	<331DEAEB.645A@pobox.oleane.com>
Date:	Wed, 05 Mar 1997 22:51:39 +0100
From:	Francois Charton <deef AT pobox DOT oleane DOT com>
Organization:	CCMSA
MIME-Version:	1.0
To:	Robert Humphris <r DOT humphris AT indigo-avs DOT com>
CC:	djgpp AT delorie DOT com
Subject:	Re: neural network code
References:	<c=GB%a=_%p=Indigo_Active_Vi%l=CRIANLARICH-970305093414Z-74 AT crianlarich DOT Indigo>

Well, sorry guys, pretty off-topic I'm afraid...

Robert Humphris wrote:
> 
> 
> Yea this is the problem, Neural nets DO NOT work! 

Hmm..., there are many industrial applications which use them. For 
example, in the US, a neural net based program (developped by AT&T) is 
used for reading zip codes on letters; a similar program (same 
developpers) has been used by some banks for reading handwritten amounts 
on checks.

Forecasting: a French water distribution company uses such a program to 
forecast water consumption/needs in Paris and its suburbs. 

Speech processing : there has been considerable experiments on use of 
Neural Nets for identifying people by the sound of their voice, or 
"denoising" some recordings, ...

> they are only designed to model the brains method of cognative function,
> not to be the brain.

Yes, but they also can be good tools for building learning machines. This 
is a common process in numerical analysis : many good algorithms were 
discovered by observing mother nature, think of simulated annealing, or 
genetic algorithms. In fact, Neural Nets have been one basis for 
investigating how statistical learning can be implemented. 

Today, the basics of why neural nets learn (in a nutshell: stochastic 
gradient algorithms and semi martingales) are well established, and new 
techniques have been discovered which yield better algorithms than neural 
nets. Hence, they got a bit out of fashion, but they are still powerful 
and easy to implement tools for machine learning.


Francois

- Raw text -