Mark F. Cook · 29 April 1992 · reply to #4013
I've got several blocks of included text, so please bear with me.
Scott Kellogg asks...
> what happens when the question of music comes up? Then of course you
> could program it to be reticent on such topics, but that's a cop out.
> It is also a tactic that could be explioted. (Blade Runner: "Tell me
> about your mother...")
>
> What do the software jocks say? James? Mark? Richard?
'Software jock'. Hmmm. I *like* that. I wonder what HP would
think if I had it added to my new business cards. :^)
In response, James writes...
> When it comes to "true" AI, existing programming models are insufficient
> to the task. The challenge to explicit programming becomes far too
> great, requiring the AIT to auto-program itself a result of life
> experiences. AIs are likely to have a very different set of initial
> experiences than humans, and would take a large amount of interaction to
> recognize and categorize human interaction and mimic it. Of course, once
> such a pattern is established in one piece of hardware, the
> implementation can be cloned, with fairly convincing results.
.... and Richard writes...
> I'll only mildly disagree with James. I suspect that it is _almost_
> within the skill of TL8 programming to create an AI that is concerto
> qualified. Of course it would do nothing else. I assert it's more a
> matter of hardware than software at this point. We can do anything,
> given enough code. :=)
I think James and Richard are both partly right. :^)
For those of you who haven't been watching it, there's a *WONDERFUL*
5-part series, sponsored in part by the ACM (that's the Association
of Computing Machinery, for those of you with non-software backgrounds),
being shown on PBS TV called "The Machine that Changed the World". As
you may have guessed by the title, the subject is digital computers. It
just so happens that this weeks episode (entitled "The Thinking Machine")
was about the attempts (and failures) to develop a true machine intelligence.
Specifically, one that could pass the Turing Test. (If you don't know
what that is, drop me a line and I'll explain it.)
According to current experts in the field, hardware *is* the limiting
factor. As early as the late-50s, there were experiments (mostly conducted
by the U.S. Army) attempting to create artificial intelligence. Later,
in the 60s, there were numerous attempts in the private sector.
They all failed. By 1980, with one exception, they were all terminated.
The only experiment still currently underway is the 'Cyke' (for
'Encyclopedia') Project in Texas. They've been constructing their
database for over ten years and they maintain that they've still got
a long ways to go.
The problem is that for years, experts insisted that you didn't have to
model the brain (hardware). All you had to do was successfully model
the 'mind' (software). The trouble is, they haven't been able to figure
out exactly how a human brain programs itself as a person grows up.
The cloest anyone has come to success falls into two catagories: expert
systems and neural networks.
Expert systems can be developed on conventional computer hardware, and
are remarkably successful, but they're confined to a single, narrowly
defined area of expertise. Outside that area, they're useless. To be
a true AI is to be able to cope with *any* subject intelligently (but
not necessarily expertly). Even the ability to gracefully admit ignorance,
in what appears to be a sentient fashion, is currently beyond the scope
of even the most ambitious AI yet constructed.
Neural networks, on the other hand, are not so much programmed, as they
are 'taught', using methods very similar to teaching a person. There
are three main problems with them. Number one, they can't be built
complex enough yet to come even close to storing the number of experiences
that a human brain can hold. Number two, even the researchers that build
them aren't exactly sure how they 'learn' the lessons they are taught.
And that leads directly to problem number three. Since it isn't clear
how a neural network 'learns', it hard to tell exactly *WHAT* it's
learning. It may not be what you think. Here's a real example.
In the early 1960s, the U.S. Army decided to build an AI (primitive
neural network) which could do high speed tactical recon photo analysis.
The goal was to have the AI rapidly identify enemy tanks in aerial
photos. To do this, the Army took over 2,000 aerial photos, all of
outdoor scenes (forested, open fields, and combinations thereof).
Half of the photos contained one or more tanks (some in plain sight,
and some partially concealed by foliage). The other half did not
contain any type of vehicle. About 90% of these photos were used to
'teach' the AI the difference between the two types of photos. When
the instruction was finished, the remaining photos were presented for
analysis. The AI scored 100% on the test photos.
Then, just to make sure that the AI was programmed correctly, the Army
went out and took more photos of both types (with and without tanks),
and presented them to the AI for interpretation. It failed almost
totally.
It turned out that in the original set of photos, all the pictures
containing tanks were taken on one day (a sunny day) and the 'empty'
pictures were taken on another day (an overcast day). All the AI
ended up learning was how to tell between nice weather and not-so-nice
weather.
End of experiment.
It would appear, based on the last 35+ years of research, that a true AI
cannot be programmed. It must learn, the same way that humans (or animals,
for that matter) learn. Even the basis data, from which learning begins,
must consist of some fairly non-intuitive information. This is what the
'Cyke' project is all about.
They discovered very early in their experiment, that the single biggest
hurdle for a budding AI to overcome is that of understanding the context
of human communication. Here are two examples.
"Jane looked through the glass window of shop at the red bicycle.
She very badly wanted it."
"Jane looked at the red bicycle through the glass window. She
pressed her nose up against it as she stared."
To humans, it is very apparent that 'it' in the first sentence refers
to the bicycle, but 'it' in the second sentence refers to the glass
window. An AI would be unable to make the distinction because of it's
lack of understanding of human context. Humans develop this ability
automatically in early childhood. A compute must be taught explicitly.
To this end, Project 'Cyke' is attempting to teach a computer human
context by brute-force. They are programming an entire encyclopedia
into the database of their AI, but they're not just including what
one would find in a normal encyclopedia. They must also include those
bits of information that, to a human, would seem so ridiculously obvious
that no one would ever think it necessary to codify them. For example,
under the encyclopedia entry for "Abraham Lincoln", they are including
data points like:
* When Lincoln became president of the United States and was
in Washington D.C., his left foot was also in Washington D.C.
* Lincoln's name and date of birth stayed the same during his
entire life.
* Lincoln's children were always younger then Lincoln.
* When Lincoln died, he remained dead, although some of his ideas
and knowledge of him continued to exist.
Sounds pretty obvious, huh? Well, not to a computer. The researchers
on the 'Cyke' Project are still many years away from testing their 'baby',
but they run it every night to have it look for inconsistancies in it's
database, to establish correlations on it's own. Some of the results
have been... strange. For example:
SetOfPersons = Preeminent
Cyke says all people are preeminent. This an artifact of the fact that,
so far, all of the individuals programmed into the database (with the
exception of the project researchers themselves) have been famous people.
Therefore *ALL* people are famous. Or how about:
SetOfNationsSameGovernmentSameReligion = SetOfNationsSameOrganizations
Cyke says all nations with the same general government type and same primary
religion are also members of the same organizations. (I.e. all moslem
monarchies are members of OPEC.) My example here is almost certainly wrong,
but Cyke's correlation may be correct. The researchers are attempting to
validate or invalidate the premise.
And the work continues.
Richard also writes...
> Actually, my view of machine intelligence is that when it arrives,
> it will in no way mimic human intelligence. Well, maybe superficially.
> After all, the robot - Si creature that it is, rugged, semi-immortal,
> with different sensor apparatus, has about as much in common with us
> as does the fruit fly.
Sad, but probably true. Thank God for PGMPs, huh? :^)
So Scott asks...
> Sounds cool, but what would *that* be like? What would it think like?
Beats me (and everybody else too, I suspect). Ask us what the color 'blue'
tastes like. That's an easier question to answer. :^)
I've got a more specific question: would 'garbage collection' routines
for an AI (should they need to exist) be analogous to human dreaming?
Actually, I'm saving my *GOOD* answers to your question for Colin Roald's
summer PBEM (I hope).
Later,
"Nothing becomes real till it is experienced -
even a proverb is no proverb to you till your
life has illustrated it."
- John Keats
"I can only assume that a 'Do Not File' document
is filed in a 'Do Not File' file."
- Senator Frank Church,
Senate Intelligence Subcommittee
Hearing, 1975
- Mark C.