John, (01)
We have a way of talking past one another. I am continuing this thread,
but with some concern that it may be pointless. (02)
> JB> It seems to me that AI has long abandoned practical application
> > areas in favor of the theoretic work currently in vogue...
> >
> > Clearly the future of AI, cognitive science or semantic processing
> > must include tight coupling with real world problems.
>
> EB> I think the scope of this statement is too grand.
> > Do we regard AI as
> > - primarily a (natural) science?
> > - primarily a formal or philosophical discipline?
> > - primarily an engineering discipline?
>
> JS: All three of those areas have been developed in detail over the past
> half century, and many AI researchers/developers have moved freely
> from one to the other and back again. (03)
Of course. My point was that the engineering concern is, or should be,
real-world problems. But it is not a requirement, or even necessarily a
good idea for the other two major aspects of AI research. You do seem
to disagree with that tenet later in your email. (04)
> EB> The development of AI technologies -- algorithms for reasoning
> > -- is an engineering discipline. Its objective is to produce
> > useful tools for reasoning to effect about real-world problems.
>
> I agree that the goal of engineering is to solve real-world problems.
> But the AI culture has been so isolated from the mainstream of
> application development that most AI researchers wouldn't recognize
> a real-world problem if they stumbled over it. (05)
Which may only be to say the "most AI researchers" work in the cognitive
science, formal logic, and formal semantics areas. And I think that is
probably true of the real "researchers". There is also a large group of
students who are engineering worthless AI tooling and calling it
"research", but that is not the point -- the value of that work is that
they are becoming educated. There are a good many software engineers
now using "AI technologies" for commercial purposes. The difference is
that they don't publish their work. (06)
> A prime example is the insanity of designing the Semantic Web
> without recognizing that every major web site is built around
> a relational database. (07)
An interesting aside, which I forfend to pursue. (08)
> JFS>> AI is dominated by brilliant people who are totally out of touch
> >> with anything and everything that goes on in the field of commercial
> >> data processing.
>
> EB> My gut reaction is: and rightly so. Most commercial data processing
> > is not very interesting. The technologies needed to do it well were
> > devised over the 30 years 1965-1995 and they are heavily and reliably
> > used.
>
> First of all, I would never say "rightly so." Too many researchers in
> every field are prima donas who are afraid to get their hands dirty.
> I have no sympathy for that attitude. (09)
I would say that there is a difference between being "afraid to get
one's hands dirty" and having the experience to know that one commercial
application in 20 needs an automotive engineer, 9 need auto mechanics
and the other 10 just need someone with a class A driver's license. (010)
I agree that there are researchers who don't want to get their hands
dirty. But there are many others who are willing and able, but tire
quickly of being asked to solve problems that many of their students
could solve. For example, you don't need AI to fix bad database
designs; you just need a competent database analyst and the
organizational will to do the redesign. (011)
(I have twice been contracted to tell an organization what its in-house
experts already knew, simply because their management didn't believe
them without confirmation from some respected source. I dare say many
senior AI folk have had similar bitter experiences.) (012)
> I'll agree that many commercial applications are not as exciting,
> but many of them, especially the larger ones, are very challenging.
> For examples, just look at some of the large projects that the US
> gov't and various large corporations pay for, but turn out to be
> very expensive failures. (013)
That is a different problem. In most cases, these people are trying to
hire first-class expertise to do the work. The problem is that the
people with the problem don't necessarily know what expertise is best
suited to solving it. The question is: How do you get the guy with the
interesting commercial problem and the AI expert who can solve it at the
same cocktail party? The commercial partner search technologies are
exploited by people whose expertise is in getting the contracts, not in
employing the best technologies for solving the problems. In fact, they
are more likely to be successful by proposing a tried-and-true approach
than a revolutionary one. (I once worked for an organization that lost
two major contracts by proposing something the evaluator didn't expect.) (014)
> AI technology was migrating to commercial applications 50 years ago.
> Just look at the contributions from LISP:
>
> Recursive functions, list processing, the if-then-else statement
> including multiple elseif's, automatic storage management and
> garbage collection, lambda expressions, functional programming,
> metalevel programming, the ability to manipulate programs, etc. (015)
Well, I don't want to debate how many of these ideas were original with
John McCarthy. What I would observe is:
1) LISP is a programming language used by AI folk. There is no
intrinsic AI quality to LISP, unless you think "metalevel programming"
(i.e., S-expressions) somehow is.
2) The number of commercial applications written in LISP in the last
45 years is very small, primarily because LISP implementations were so
self-contained that getting them to read commercial data from things
like tapes and databases did not become common for 20 years. (016)
And you could probably count on the fingers of one hand the number of
commercial applications that were written in SAIL. (017)
Prolog, OTOH, has had some real commercial successes, and datalogic and
its relatives have been widely used in the last 20 years. And the last
15 years or so has seen dramatic growth in the use of rules engines,
because the hardware enabled it to be useful and the tools were linked
to databases and popular programming languages. (018)
At the same time, I would say that declarative programming and
condition/action languages are not exactly cutting-edge AI. The AI
technologies of 1960-1990 have become commercially more commonplace, but
they are no longer the domain of AI research. Which is only to say
that, as in other disciplines, the new ideas migrate into the education
curricula over 10 years, and harden into useful engineering tooling and
practice, and 5-10 years after that they become common in industry. (019)
So, what the great minds of AI are doing now will be taught to
undergraduate computer scientists in 2020 and be common in industry
around 2025. We might be able to shorten that cycle a bit, but the
pattern is well-established. (And all of that, BTW, was encapsulated in
"and rightly so". ;-)) (020)
> For many years, I was telling Doug Lenat that he should devote more
> attention to implementing applications. But he said that he didn't
> want to dilute the "pure" research by diverting resources to
> applications. (021)
I am going to say "dicto simpliciter". I don't deny that that attitude
is common in ivory towers, and if, like Doug, you are paid millions to
do "pure" research, your tower will stand firm. But I don't think you
will find that that attitude is common on this exploder, for example. (022)
> I believe that attitude was counterproductive both for Cyc and
> for the broader field of AI. (023)
And you are probably right. But the whole Cyc project is very atypical
of AI research from 1970 to 2000. It was better funded by orders of
magnitude, and by a Defense organization with motives that precluded
wider education in, and use of, the Cyc technologies. (024)
Similarly, NSA has long had better search engines than Google and its
new rivals, I don't doubt. But it is not because of the attitudes of
people doing research in information retrieval on the Web. (025)
> Second, doing research without any clear idea of how it is going
> to be used is a recipe for creating products with no clear use. (026)
Well, I doubt that most scientists think what they do is a recipe for
creating products of any kind. Commercial scientific research tends to
be directed into areas in which the knowledge discovered is expected to
be valuable to their commercial interests. And in consequence, the
knowledge discovered is kept secret. We should thank our lucky stars
that a large portion of scientific research is not so encumbered. (027)
> Third, pure mathematics has benefited enormously from applications,
> ranging from surveying land in Egypt to modern physics. Without
> that stimulus, mathematics would be in a primitive state. (028)
Mathematicians are intrigued by problems. At least half of those
problems probably did originate in real-world problems. We all know
about the bridges of St. Petersburg, but I'm not sure it was a practical
problem. I never heard what the practical significance of solving
quintic equations was, or why anyone cared about Fermat's Last Theorem
(which, I was told, gave rise to category theory). And I have no idea
what real-world problems inspired Frege and Russell, if any. So I'm not
willing to accept your last sentence as generally true. (029)
> EB> An EU study ending in 2007 concluded that we now have a lot
> > of AI tooling, but we don't have much encoded knowledge.
>
> I disagree with that study. At VivoMind, we have been getting
> excellent results from automated and semiautomated methods
> for *learning* the knowledge. I believe that Cyc would have
> discovered that point years ago, if they had worked on real
> applications. (030)
(I was bit sloppy in what I said. I think there is an enormous amount
of knowledge that is now encoded in some form: databases, PDF, HTML,
etc. What was meant was "knowledge encoded in a form suitable for
reasoning".) (031)
I don't see how your immediate experience in capturing knowledge has
much to do with the amount of encoded knowledge available for using AI
applications to solve real-world problems. Are there hundreds of
organizations doing what you are doing? Is your technology being
rapidly proliferated as we speak? If so, then what you are saying is
not that the 2007 study was wrong, but that that situation will now
change very rapidly. Is that your point? (032)
-Ed (033)
--
Edward J. Barkmeyer Email: edbark@xxxxxxxx
National Institute of Standards & Technology
Manufacturing Systems Integration Division
100 Bureau Drive, Stop 8263 Tel: +1 301-975-3528
Gaithersburg, MD 20899-8263 FAX: +1 301-975-4694 (034)
"The opinions expressed above do not reflect consensus of NIST,
and have not been reviewed by any Government authority." (035)
_________________________________________________________________
Message Archives: http://ontolog.cim3.net/forum/ontolog-forum/
Config Subscr: http://ontolog.cim3.net/mailman/listinfo/ontolog-forum/
Unsubscribe: mailto:ontolog-forum-leave@xxxxxxxxxxxxxxxx
Shared Files: http://ontolog.cim3.net/file/
Community Wiki: http://ontolog.cim3.net/wiki/
To join: http://ontolog.cim3.net/cgi-bin/wiki.pl?WikiHomePage#nid1J
To Post: mailto:ontolog-forum@xxxxxxxxxxxxxxxx (036)
|