[Top] [All Lists]

[ontolog-forum] More about the IBM Watson project

To: "'[ontolog-forum] '" <ontolog-forum@xxxxxxxxxxxxxxxx>
From: John F Sowa <sowa@xxxxxxxxxxx>
Date: Tue, 18 Sep 2012 09:03:56 -0400
Message-id: <5058713C.3010109@xxxxxxxxxxx>
There have been many short articles about Watson that discussed some
details, but they didn't give a clear picture about how the details fit
together.  I recently came across a one-hour talk by Chris Welty that
contained more technical material about the overall organization:    (01)

    http://videolectures.net/eswc2012_welty_watson/    (02)

The first 12 minutes give a rehash about the Jeopardy game.  You can
skip that by clicking on the list of slides to the right of the video
screen.  Start the video and click on the following line:    (03)

    12:11 Welty's Trident    (04)

I have some qualifications about what Welty says in that slide,
but I strongly agree with his point that you need to have multiple
paradigms and a learning method for choosing among the possibly
conflicting information that comes back.  The crucial discussion
is about the following slide:    (05)

   15:12 A new software paradigm is emerging    (06)

In the next 20 minutes, Welty presents the details about how
Watson generates multiple hypotheses about possible answers,
evaluates each answer on a large number of dimensions (features),
and uses a perceptron algorithm to learn how to weight the
features to get a combined measure of confidence in the answer.    (07)

For comparison, I'd like to mention an article written by
Arun Majumdar and me, entitled "Two Paradigms Are Better Than One,
and Multiple Paradigms Are Even Better" (abstract and URL below).    (08)

That article cites Marvin Minsky's _Society of Mind_ as an
inspiration.  I agree with Welty on those points where he agrees
with Minsky:  the need for heterogeneous methods, some way to
evaluate the results of different methods, and learning methods
to determine which one(s) are appropriate to a particular task.    (09)

But where I disagree with Welty is on the level of granularity.
Minsky emphasizes that the Mind (or brain) has heterogeneous modules
that operate at every level of granularity.  But the Watson learning
methods are applied at much larger granules.  In the article below,
we emphasize the use of heterogeneous methods for analyzing and
interpreting language at every level of detail.    (010)

For further discussion of related issues, see the compendium
of slides in http://www.jfsowa.com/talks/goal.pdf    (011)

_________________________________________________________________    (012)

Source: http://www.jfsowa.com/pubs/paradigm.pdf    (013)

                 Two Paradigms Are Better Than One,
               And Multiple Paradigms Are Even Better    (014)

                 Arun K. Majumdar and John F. Sowa    (015)

Abstract. During the past half century, the field of artificial
intelligence has developed a large number of theories, paradigms,
technologies, and tools. Many AI systems are based on one dominant
paradigm with a few subsidiary modules for handling exceptions or
special cases. Some systems are built from components that perform
different tasks, but each component is based on a single paradigm.
Since people freely switch from one method of thinking or reasoning
to another, some cognitive scientists believe that the ability to
integrate multiple methods of reasoning is key to human-like
flexibility.    (016)

In his book _The Society of Mind_, Minsky (1986) presented an
architecture for intelligence based on a society of heterogeneous
agents that use different reasoning methods to solve different
problems or different aspects of the same problem. That idea is
intriguing, but it raises many serious issues: how to coordinate
multiple agents, distribute tasks among them, evaluate their results,
encourage agents that consistently produce good results, inhibit
agents that produce misleading, irrelevant, or unfruitful results,
and integrate all the results into a coherent response. The most
difficult problem is to enable multiple heterogeneous agents, acting
independently, to produce the effect of a single mind with a unified
personality that can pursue and accomplish coherent goals. This article
discusses ways of organizing a society of heterogeneous agents as an
integrated system with flexible methods of reasoning, learning, and
language processing.    (017)

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    (018)

<Prev in Thread] Current Thread [Next in Thread>