TITLE: AI's Biggest Challenges Are Still To Come
AUTHOR: Eugene Wallingford
DATE: May 27, 2018 10:20 AM
A lot of people I know have been discussing
last week's NY Times op-ed
about recent advances in neural networks and what they mean for
AI. The article even sparked conversation among colleagues from
my grad school research lab and among my PhD advisor's colleagues
from when he was in grad school. It seems that many of us are
frequently asked by non-CS folks what we think about recent
advances in AI, from AlphaGo to voice recognition to self-driving
cars. My answers sound similar to what some of my old friends
say. Are we now afraid of AI being able to take over the world?
Um, no. Do you think that the goals of AI are finally within
reach? No. Much remains to be done.
I rate my personal interest in recent deep learning advances as
meh. I'm not as down on the current work as the authors
of the Times piece seem to be; I'm just not all that
interested. It's wonderful as an exercise in engineering:
building focused systems that solve a single problem. But, as
the article points out, that's the key. These systems work in
limited domains, to solve limited problems. When I want one of
these problems to be solved, I am thankful that people have
figured out how to solve and make it commercially available for
us to use. Self-driving cars, for instance, have the potential
to make the world safer and to improve the quality of my own life.
My interest in AI, though, has always been at a higher level:
understanding how intelligence works. Are there general
principles that govern intelligent behavior, independent of
hardware or implementation? One of the first things to attract
me to AI was the idea of writing a program that could play chess.
That's an engineering problem in a very narrow domain. But I
soon found myself drawn to cognitive issues: problem-solving
strategies, reflection, explanation, conversation, symbolic
reasoning. Cognitive psychology was one of my favorite courses
in grad school in large part because it tried to connect
low-level behaviors in the human brain connected to the symbolic
level. AlphaGo is exceedingly cool as a game player, but
it can't talk to me
about Go, and for me that's a lot of the fun of playing.
In an email message earlier this week, my quick take on all this
work was: We've forgotten
the knowledge level.
And for me, the knowledge level is what's most interesting about
That one-liner oversimplifies things, as most one-liners do. The
AI world hasn't forgotten the knowledge level so much as moved
away from it for a while in order to capitalize on advances in
math and processing power. The results have been some impressive
computer systems. I do hope that the pendulum swings back soon
as AI researchers step back from these achievements and builds
some theories at the knowledge level. I understand that this may
not be possible, but I'm not ready to give up on the dream yet.