Seattle — Inside the Allen Institute for Artificial Intelligence,
known as AI2, everything is a gleaming architectural white. The walls
are white, the furniture is white, the counters are white. It might as
well have been a set for the space station in “2001: A Space Odyssey.”
“The
brilliant white was a conscious choice meant to evoke experimental
science — think ‘white lab coat,’ ” said Oren Etzioni, a computer
scientist and director of the new institute, which the Microsoft
co-founder Paul Allen launched this year as a sibling of the Allen Institute for Brain Science, his effort to map the human brain.
Yet
for the 30 (soon to be 50) artificial-intelligence researchers who can
look out on a striking view of downtown Seattle, the futuristic
surroundings offer a paradoxical note: AI2 is an effort to advance artificial intelligence while simultaneously reaching back into the field’s past.
While Silicon Valley looks to fashionable techniques like neural networks and machine learning
that have rapidly advanced the state of the art, Dr. Etzioni remains a
practitioner of a modern version of what used to be known as Gofai, for
good old-fashioned artificial intelligence.
The
reference goes back to the earliest days of the field in the 1950s and
’60s, when artificial-intelligence researchers were confident they could
model human intelligence using symbolic systems — logic embedded in
software programs, running on powerful computers.
Then
in the late 1980s, an early wave of commercial artificial-intelligence
companies failed, bringing on what became known as the “A.I. winter.”
The field was seen as a failure and went into eclipse.
In
recent years, however, A.I. has come roaring back as speech
recognition, machine vision and self-driving cars have made progress
with powerful computers, cheap sensors and machine-learning techniques.
That has started a Silicon Valley gold rush led by Google, Facebook and
Apple, drawing outsiders like Alibaba and Baidu in China, all caught up
in a frantic race to hire the world’s best machine-learning talent.
But
the debate over how to reach genuine artificial intelligence has not
ended, and Dr. Etzioni and Mr. Allen are betting that their path is more
pragmatic. The power of the new techniques is not disputed, but there
is a growing debate over whether they can take the field to human-level
capabilities by themselves.
“Think
of it as Sherlock Holmes versus Spider-Man,” said Jerry Kaplan, a
visiting lecturer at Stanford who teaches a course on the history and
philosophy of artificial intelligence, comparing Holmes’s deductive
powers with the irrational “spider sense” that tingles at the base of
Spider-Man’s skull and alerts him to danger.
Mr.
Allen, who noted that he came from a family of librarians, said his
decision to fund an artificial-intelligence research lab was inspired by
the question of how books and other knowledge might be encoded to
become the basis for computer interactions in which human questions
might be answered more fully.
Continue reading the main story
“AI2
was born from a desire to create a system that could truly reason about
knowledge, rather than just offer up what had been written on a subject
before,” he wrote in an email interview.
Dr.
Etzioni says that the artificial-intelligence field has made
incremental advances in areas like vision and speech, but that we have
gotten no closer to the larger goal of true human-level systems.
“Driverless
cars are a great thing,” he said, but added that the field had given
rise to “bad A.I., like the N.S.A. is using it or Facebook is using it
to track you.”
“We want to be the good guys,” he went on, “and it’s up to us to deliver on that.”
Moreover,
he says, both he and Mr. Allen believe that technology cannot be
separated from its social and economic consequences. They have added a
social mission to the project that they call “artificial intelligence
for the common good.”
The
success or failure of the project, however, will ultimately hinge on
whether Dr. Etzioni can create a new synthesis of artificial
intelligence, weaving together powerful machine-learning tools with
traditional logic-oriented software.
The
current fad for big data, of which machine learning is a major
component, has significant limits. “If you step back a little and say we
want to do A.I., then you will realize that A.I. needs knowledge,
reasoning and explanation,” he said. “My argument is that big data has
made great progress in limited areas.”
Even Watson,
the brainy IBM computer whose intelligence the company wants to apply
in complex applications like medical diagnoses and automated call
centers with interactive speech recognition, will soon reach fundamental
limits, he argues.
“I
really don’t want a system that can’t explain itself to be my doctor,”
he said. “I can just imagine sitting there with Dr. Watson and the
program saying, ‘Well, we need to remove a kidney, Mr. Etzioni,’ and I’m
like, ‘What?!’ and they respond, ‘Well, we have a lot of variables and a
lot of data, and that’s just what the model says.’ ”
Dr.
Etzioni, 50, was already known for innovative web projects, including
MetaCrawler, an early search engine, and an array of successful start-up
companies; one of them, Farecast, was acquired by Microsoft and became
the basis for its Bing Travel service. (The first student to major in computer science at Harvard, he is a son of the well-known sociologist Amitai Etzioni.)
At
AI2 he is motivated by Mr. Allen’s view that “in order to be truly
intelligent, computers must understand — that is probably the critical
word,” as the Microsoft co-founder put it in a 1977 interview.
Some
technology experts argue that self-aware computing machines are now on
the horizon. “As for A.I. progress, we’re mostly haggling about a few
decades,” said Hans Moravec, a leading roboticist who is the chief
scientist of Seegrid Corporation, a maker of autonomous vehicles for
warehouse applications. “I’m content to simply watch it play out, trying
to do my part. I do want fully autonomous robots as soon as possible,
to begin visiting the rest of the universe.”
Continue reading the main story
Continue reading the main story
Continue reading the main story
Mr.
Allen and Dr. Etzioni are not so optimistic. Both are skeptical of
claims that we may be only years away from machines that think in any
human sense.
“Full
A.I., in the sense of something like HAL in ‘2001,’ ” Mr. Allen wrote
in an email interview, “is probably a hundred years away (or more). In
reality, we are only beginning to grasp how deep intelligence works.”
Dr.
Etzioni wants AI2 to set measurable goals to help get a new class of
learning systems off the ground. During its first year, the researchers
have focused on three projects — one in computer vision
(in which computers learn to recognize images), one to build a
reasoning system capable of taking standardized school tests, and a
third to help scholars deal with the fire hose of information that is
inundating every scientific field.
The school-test effort, Project Aristo,
seeks to create a learning program that can collect and organize a wide
range of information, and then use that database to reason and to
answer questions, even discussing and explaining its answers with human
users.
To
chart Aristo’s progress, researchers plan to test it on increasingly
difficult standardized science exams, moving from the fourth grade
through the 12th.
“We’re
not planning on putting 10th graders out of work,” Dr. Etzioni said.
But he does believe that a program that can converse with humans and
answer questions would serve as a foundation for many other
achievements, going far beyond the most powerful search engines and
systems like Watson.
In
September, the researchers celebrated their first milestone — 60
percent correct answers in the language portion of New York State’s
fourth-grade science test. Many of the questions in the actual test
include diagrams and illustrations, which will ultimately require
advances in computer vision.
That
challenge is considered far more difficult than recognizing human
speech. It calls for a computer system with “scene understanding,” the
human ability to extract meaning from animate and inanimate objects that
interact.
Whether
AI2’s research leads to a new generation of thinking machine or just
more incremental advances, the project is a clear indication that
artificial intelligence has once again become the defining force in the
software world.
“The
narrative has changed,” said Peter Norvig, Google’s director of
research. “It has switched from, ‘Isn’t it terrible that artificial
intelligence is a failure?’ to ‘Isn’t it terrible that A.I. is a
success?’ ”
No comments: