From: Andy Wilson <ajw@Think.COM>
Date: Thu, 26 Aug 93 12:22:32 PDT
To: bill@twwells.com
Subject: Visa, HNC Inc. develop neural network as a weapon to fight fraud
In-Reply-To: <CCCEF6.9rz@twwells.com>
Message-ID: <9308261917.AA16210@custard.think.com>
MIME-Version: 1.0
Content-Type: text/plain


   From: bill@twwells.com (T. William Wells)
   Date: Thu, 26 Aug 1993 01:21:03 GMT

   [...]

   But that wasn't what you were writing about. You were writing
   about bad business decisions, not violations of privacy.

No, you were writing about bad business decisions.  I was providing
a few details on how credit/charge-card information is used in this
process and a few potential problems resulting from it.

   For that matter, your notions on neural networks seem
   contradictory. On the one hand, you complain about a violation of
   privacy and on the other you complain that a neural network won't
   tell you how it reached its conclusions!

You are deliberately confusing two different points: 1) the fact that
neural networks do not provide useful explanations of how they arrived
at a particular decision,  and 2) some potential problems that arise
from this fact that concern privacy issues.

   :                I beg to differ.  This is exactly what digital
   : cash is meant to prevent.

   Digital cash and the use of neural networks to authenticate
   transactions are essentially orthogonal issues.

I will reiterate that the whole point of digital cash is to provide
anonymity,  which will prevent these kinds of uses made of personal
information which are not done with the explicit approval of the
person involved.

   : The problem with referring a neural network's decision to a human
   : is that the neural network gives no information other than the
   : probability of fraud.

   1) This statement is false. It is true of some neural networks
      but not all. We have no way of knowing whether their neural
      network is among those.

It is true of all commercial applications of neural networks to my
knowledge,  and certainly true of the neural networks developed
by Hecht-Nielsen.

   :                                                     There is not any
   : good way to combine the judgement of the neural net with that of a
   : human for that reason.

   Nonsense. As the existence of rule based systems that incorporate
   neural networks shows.

That shows no such thing.  The only way to combine the judgement of
a neural network with that of a rule-based system,  or anything else,
is to see if both arrive at the same conclusion.  You cannot see the
reasoning process of the neural network to help the human understand
why it made the judgement that it did,  the marketing hype of neural
network vendors notwithstanding.

This is my last post on this thread.

Andy