Two notes. (01)
John F. Sowa wrote:
> Tonight on PBS, there is a program about the IBM computer that plays
> Jeopardy. The following IBM site lists the names of the participants
> on the project:
>
> http://www-943.ibm.com/innovation/us/watson/research-team/algorithms.html
>
> I know a few of the people from my olden days at IBM, but most
> of them joined after I left IBM
>
> Their project leader, David Ferrucci, gave a lecture, which is
> available on the IBM web site:
>
>
>http://www-943.ibm.com/innovation/us/watson/watson-for-a-smarter-planet/building-a-jeopardy-champion/how-watson-works.html
>
> From what I gather, they are using familiar NLP techniques, but putting
> a huge amount of computer power behind them. I would call it a good
> engineering solution to a specific problem. It can probably be
> generalized to a wider range of problems.
> (02)
The 'huge amount of computer power' is supported by a huge knowledge
base, mostly in natural language, with some important markup and
indexing. The reasoning technique involves multiple parallel searches,
using a combination of associative, statistical and logical reasoning,
combined with a voting mechanism that chooses the 'most likely answer'.
The power of the reasoning approaches and the multicomputer platform
would be useless without the enormous knowledge repository. (03)
> In any case, I am very pleased to see that IBM is showing the world
> that the traditional technology of AI and NLP can really be made to
> work when you put some effort into doing it.
> (04)
That is, when you are willing to invest in 5-10 smart people for 5+
years without seeing a target product. Industry rarely has the luxury,
and universities can't pay the talented people enough to keep a
functional cadre for that long. Watson Research Center has been an
incredible place for 50 years, while many of its erstwhile competitors
(e.g., AT&T Bell Labs, Xerox PARC) have been reduced or destroyed by
industry economics. (And OBTW, it does help to have Columbia or
Princeton or Stanford just down the road.) You have to respect IBM for
its continued commitment to heavy-duty research. (05)
> IBM is doing that with a brute force supercomputer. But with better
> algorithms, it's possible to get comparable performance with much
> less hardware.
> (06)
There, John, I think you miss one of the ideas that IBM got right a long
time ago -- you have to size the computational power to the data flow
power. 'Comparable performance' in a narrow field with a much smaller
knowledge repository may be possible with much less hardware, but other
AI technologies can also be very successful in such cases. Comparable
performance in a wider-ranging field with a lot of
not-clearly-interrelated information will require a large knowledge
repository and the power to process it. You are not going to process a
terabyte knowledge base, or even a 20GB knowledge base, in 60 seconds
with your twin 4GHz desktop processors. (07)
-Ed (08)
> John
>
> _________________________________________________________________
> 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
>
> (09)
--
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 Cel: +1 240-672-5800 (010)
"The opinions expressed above do not reflect consensus of NIST,
and have not been reviewed by any Government authority." (011)
_________________________________________________________________
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 (012)
|