TITLE: The Dawn of a New Age
AUTHOR: Eugene Wallingford
DATE: February 18, 2011 4:51 PM
DESC:
-----
BODY:
My world this week has been agog with the Watson match
on Jeopardy!. The victory by a computer over the
game's two greatest champions may well signal a new
era, in much the same way as the rise of the web and
the advent of search. I'd like to collect my thoughts
before writing anything detailed about the match
itself.
In anticipation of match, last week my Intelligent
Systems students and I began to talk about some of the
techniques that were likely being used by Watson under
the hood. When I realized how little they had learned
in their AI course about reasoning in the face of
uncertainty, I dusted off an old lecture from my AI
course, circa 2001. With a few tweaks and additions,
it held up well. Among its topics were probability and
Bayes' Law. In many ways, this material is more timely
today than it was then.
Early in the week, as I began to think about what this
match would mean for computing and for the world, I was
reminded by Peter Norvig's
The Machine Age
that, in so many ways, the Jeopardy! extravaganza
heralds a change already in progress. If you haven't
read it yet, you should.
The shift from classical AI to the data-driven AI that
underlies the advances Norvig lists happened while I
was in graduate school. I saw glimpses of it at
conferences on expert systems in finance and
accounting, where the idea of mining reams of data
seemed to promise new avenues for decision makers in
business. The data might be audited financial
statements of public corporations or, more tantalizing,
collected from grocery store scanners. But I was
embedded pretty deeply in a particular way of thinking
about AI, and I missed the paradigm shift.
What is most ironic for me is that my own work, which
involved writing programs that could construct legal
arguments using both functional knowledge of a domain
and case law, was most limited by the problem that we
see addressed in systems like Google and Watson: the
ability to work with the staggering volume of text that
makes up our case law. Today's statistical techniques
for processing language make extending my work and
seeing how well it works in large domains possible in
a way I could only dream of. And I never even
dreamed of doing it myself,
such was my interest in classical AI. (I
just needed a programmer,
right?)
There is no question that data-driven, statistical AI
has proven to be our most successful way to build
intelligent systems, particularly those of a certain
scale. As an engineering approach, it has won. It
may well turn out to be the best scientific approach
to understanding intelligence as well, but... For an
old classicist like me, there is something missing.
Google is more idiot savant than bon
vivant; more Rain Man than Man on the Street.
There was a telling scene in one of the many short
documentary films about Watson in which the lead
scientist on the project said something like, "People
ask me if I know why Watson gave the answer it did.
I don't how it got the answer right. I don't know how
it got the answer wrong." His point was that Watson
is so complex and so data-rich it can surprise us with
its answers. This is true, of course, and an important
revelation to many people who think that, because a
computer can only do what we program it to do, it can
never surprise or create.
But immediately I was thinking about another sense of
that response. When Watson asks "What is Toronto? in
a Final Jeopardy! on "U.S. Cities", I want to ask it,
"What in the world were you thinking?" If it were a
human player, it might be able to tell about its
reasoning process. I'd be able to learn from what it
did right or wrong. But if I ask most computer programs
based on statistical computations over large data sets
"Why?" I can't get much more than the ranked lists of
candidates we saw at the bottom of the screen on
Jeopardy!
There is still a romantic part of me that wants to
understand what it means to think and reason at a
conscious level. Perhaps my introspection misleads me
with explanations constructed post hoc, but it
sure seems like I am able to think at a level above my
neurons firing. That feeling is especially strong when
I perform more complex tasks, such as writing a program
or struggling to understand a new idea.
So, when it comes to the science of AI, I still hope
for more. Maybe our statistical systems will become
complex and data-rich enough that they will be able
to explain their reasoning in a meaningful way. It's
also possible that my desire is nothing more than a
form of chauvinism for my own species, that the desire
to sit around and talk about stuff, including how and
why we think the way we do, is a quaint feature peculiar
to humans. I don't know the answer to this question,
but I can't shake the deep belief in an architecture
of thought and intelligent behavior that accounts for
metacognition.
In any case, it was fun to watch Watson put on such an
impressive show!
~~~~
If you are interested in such things, you might want to
(re-)read Newell and Simon's 1975 Turing Award lecture,
Computer Science as Empirical Inquiry: Symbols and Search.
It's a classic of the golden era of AI.
-----