Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!bb3.andrew.cmu.edu!nntp.sei.cmu.edu!news.psc.edu!hudson.lm.com!godot.cc.duq.edu!news.duke.edu!news.mathworks.com!tank.news.pipex.net!pipex!howland.reston.ans.net!usc!crash!apagani.cts.com!user
From: dean_abbott@partech.com (Dean Abbott)
Subject: Re: Searching for NN reference
Organization: PAR Government Systems Corp.
Date: Thu, 7 Sep 1995 20:44:29 GMT
Message-ID: <dean_abbott-0709951344290001@apagani.cts.com>
References: <DEHHAB.A3w@cee.hw.ac.uk>
Sender: news@crash.cts.com (news subsystem)
Nntp-Posting-Host: apagani.cts.com
Lines: 48

In article <DEHHAB.A3w@cee.hw.ac.uk>, ceemw@cee.hw.ac.uk (Marcus John
Williams) wrote:

> Hi there,
> 
> I was wondering whether anyone had heard the story about the neural
> net that was implemented by the American Dept of Defense (??) that
> attempted to classify photos of tanks as either russian or american.
> Without going into the story too much, the outcome was a system that
> classified tanks correctly (so the story goes). When the Berlin wall
> came down and nice glossy photos of russian tanks were available somebody
> dug up the network and tried it out and it failed miserably. Apparently
> all the network had done was learn that american tank photos had good
> lighting and russian photos (taken from spy planes?) were poor quality.
> 
> Anyhow, the long and short of it is I want to use this example in my
> dissertation and need a reference for it. Any ideas as to where I could
> find one - or even whether this story is true or not? Replies by email
> please.
> 
> Thanks,


I have seen this study described on a PBS show called "The Machine that
Changed the World", a five-part series on computers. There was one part on
artificial intelligence and neural networks where a study like the one
above was described, although the Berlin wall part of it was not
discussed. What they said was that a neural network was trained to
identify tanks in clutter (trees/bushes, etc) and it was found to do quite
well on training data, but when they tested it on real data it failed
miserably. Further investigation found that the reason was that all tanks
were in cloud cover, and all scenes of just clutter were in sunny weather
(or vica versa). In other words, the network learned to classify sunny
from cloudy scenes!

The conclusion though was unfair, I thought, as the network did fine, it
was the people that were responsible for collecting REPRESENTATIVE data
that failed. I still have the video, and PBS may be able to send a
transcript still.

Dean

-- 
Dean Abbott                      |
PAR Government Systems Corp.     |
1010 Prospect St., Suite 200     |
La Jolla, CA 92037               | 
dean_abbott@partech.com          |
