Re: [FT] Unpredictable AI

From: "Derk Groeneveld" <derk@cistron.nl>

> an interesting example is the DARPA project (I think) for armour
> recognition. Having fed the system with many pictures of bad-guy
armour
> and good-guy armour, and flagging each as good or bad-guy, the system
> could flawlessly distinguish within this set.
>
> Then, new pictures were fed, and the system messed up completely.
Turned
> out it had figured a flawless differentiation algorythm. All the bad
guy
> armour in the first set had been fotographed in the late afternoon,
> whereas the good-guy armour had been fotographed around noon. So, the
> easiest differentiation mechanism was by analysing the shadows. Sun
> overhead -> good guy. Sun low -> bad guy.
>
> Now, I'm sure I messed up some detail with respect to the original
story,
> but that's pretty much what it amounted to.

I can confirm the story is essentially correct. The perceptron
(definition at http://www.cs.bgu.ac.il/~omri/Perceptron/ )
 ended up recognising sunlit trees  vs shadowed ones.

What you may not know was the time this occurred: circa 1960.

Stanford Research Labs had a large neural-network research team in the
late 50s and early 60s, but Marvin Minsky pooh-poohed Neural Networks
in favour of Von Neumann architectures, so the whole area stalled for 25
years.
(See Chronology at http://www.calculemus.org/x/mchron.htm  and
http://www.danielnewman.com/final/history.html )

Their main successes were in the photo recognition of the large
strips of film taken by the ultra-highly-classified balloon-borne
cameras
the US was sending over the USSR at the time. You had to take a LOT
of film of essentially a random strip of the USSR, and no human
operators
could look at the 1000s of km of images without going nuts. But even a
really really basic perceptron was really good at picking out runways,
missile sites etc.

How do I know all this? My uncle, also A.E.Brain, was on the team. Which
I only learnt about when doing some research for my own military AI...
My Uncle, unlike nearly everyone else in AI, does not think that Marvin
Minsky's Posterior is a source of Solar Energy. I tend to agree.

 I've - not written - not made - not grown, though that's closest - um,
been
responsible for the creation of an AI system for anti-missile defence.
Basically a simple reflexive one, that evolved using genetic algorithms
 (though they weren't called that in 1994...).

My take on the subject? Confine the problem domain tightly enough,
and the AI can be made arbitrarily intelligent. Slacken the problem
domain
a bit, and its IQ drops precipitously. We're an awfully long way away
from
high-level (human,chimp, dog ) intelligence. I believe we'll get there
in
the
end, but the more we know about the subject, the more we realise that
Intelligence is not manufactured, it evolves. Rule-based systems - like
the one I caused to be created - are, if not a dead-end, just a useful
tool to be used within a controlling Neural Net. Conscious vs
Subconscious
processing.

That's why I like a Human controlling a bunch of specialised AIs. Far
better