The concept of nueral networks was behind the development of Artificial Intelligence (AI)...
The Godfathers of the AI Boom Win Computing’s Highest Honor
(from left to right) Yann LeCun, Geoff Hinton, and Yoshua Bengio |
In the late 1980s, Canadian master’s
student Yoshua Bengio became captivated by an unfashionable idea. A
handful of artificial intelligence researchers were trying to craft
software that loosely mimicked how networks of neurons process data in
the brain, despite scant evidence it would work. “I fell in love with
the idea that we could both understand the principles of how the brain
works and also construct AI,” says Bengio, now a professor at the
University of Montreal.
More than 20 years later, the tech
industry fell in love with that idea too. Neural networks are behind the
recent bloom of progress in AI that has enabled projects such as self-driving cars and phone bots practically indistinguishable from people.
On
Wednesday, Bengio, 55, and two other protagonists of that revolution
won the highest honor in computer science, the ACM Turing Award, known
as the Nobel Prize of computing. The other winners are Google researcher
Geoff Hinton, 71, and NYU professor and Facebook chief AI scientist Yann LeCun, 58, who wrote some of the papers that seduced Bengio into working on neural networks.
The
trio’s journey is a parable of scientific grit and a case study in the
economic value of new forms of computing. Through decades of careful
research out of the limelight, they transformed an old-fashioned,
marginalized idea into the hottest thing in computer science. The
technology they championed is central to every large tech company’s
strategy for the future. It’s how software in testing at Google reads medical scans, how Tesla’s Autopilot reads road markings, and how Facebook automatically removes some hate speech.
Asked
what winning the Turing Award means, Hinton expresses mock surprise. “I
guess neural networks are now respectable computer science,” he says.
The joke is that in computer science, there isn’t anything more
respectable than a Turing Award. It has been awarded annually since 1966
and is named after Alan Turing, the British mathematician who laid some of the early foundations for computing and AI in the 1930s, ’40s, and ’50s.
Pedros
Domingos, a professor at the University of Washington who leads machine
learning research at hedge fund DE Shaw, says it’s beyond time that
deep learning was recognized. “This was long overdue,” he says.
Domingos’ 2015 book The Master Algorithm surveyed five “tribes” taking different approaches to AI, including the “connectionists” working on neural networks.
Awarding
the Turing to that tribe acknowledges a shift in how computer
scientists solve problems, he says. “This is not just a Turing Award for
these particular people. It’s recognition that machine learning has
become a central field in computer science,” says Domingos.
The
discipline has a long tradition of valuing mathematically proven
solutions for problems. But machine learning algorithms get things done
in a messier way, following statistical trails in data to find methods
that work well in practice, even if it’s not clear exactly how. “Computer science is a form of engineering, and what really matters is whether you get results,” Domingos says.
The idea of a “neural network” is one of the oldest approaches to artificial intelligence,
dating back to the emergence of the field in the late 1950s.
Researchers adapted simple models of brain cells created by
neuroscientists into mathematical networks that could learn to sort data
into categories by filtering it through a series of simple nodes, which
were likened (rather superficially) to neurons.
Early successes included the room-filling Perceptron, which could learn to distinguish shapes on a screen. But it was unclear how to train large networks with many layers of neurons, to allow the technique to go beyond toy tasks.
Hinton showed the solution to training so-called deep networks. He coauthored a seminal 1986 paper
on a learning algorithm called back-propagation. That algorithm, known
as backprop, is at the heart of deep learning today, but back then the
technology wouldn’t quite come together. “There was a blackout period
between the mid-’90s and the mid-2000s where essentially nobody but a
few crazy people like us were working on neural nets,” says LeCun.
His contributions included convnets, invented neural
network designs well suited to images; he proved the concept by creating
check-reading software for ATMs at Bell Labs. Bengio pioneered methods
to apply deep learning to sequences, such as speech, and understanding
text. But the wider world only caught on to deep learning early in this
decade, after researchers figured out how to harness the power of
graphics processors, or GPUs.
One crucial moment
took place in 2012, when Hinton, then at the University of Toronto, and
two grad students surprisingly won an annual contest for software that
identifies objects in photos. Their triumph left the field’s favored
methods in the dust, correctly sorting more than 100,000 photos into
1,000 categories within five guesses with 85 percent accuracy, more than
10 percentage points better than the runner-up. Google acquired
a startup founded by the trio early in 2013, and Hinton has worked for
the company ever since. Facebook hired LeCun later that year.
“You
can look back on what happened and think science worked the way it's
meant to work,” Hinton says. That is, “until we could produce results
that were clearly better than the current state of the art, people were
very skeptical.”
Hinton says he and his
collaborators stuck with their unfashionable ideas for so long because
they are mavericks at heart. All three are now part of the academic and
tech industry mainstream. Hinton and LeCun are vice presidents at two of
the world’s most influential companies. Bengio has not joined a tech
giant, but he is an adviser to Microsoft and has worked with startups
adapting deep learning to tasks such as drug discovery and helping
victims of sexual harassment.
The three have gone
in different directions, but they remain collaborators and friends.
Asked whether they will deliver the traditional Turing Award lecture
together, Hinton raises chuckles by suggesting Bengio and LeCun go first
so he can give his own lecture about what they got wrong. Does that
joke reflect the trio’s typical working dynamic? Hinton says no at the
same time LeCun good-naturedly answers yes.
Despite
deep learning’s many practical successes, there’s still much it can’t
do. Neural networks are brain-inspired but not much like the brain. The
intelligence that deep learning gives computers can be exceptional at
narrowly defined tasks—play this particular game, recognize these
particular sounds—but isn’t adaptable and versatile like human
intelligence.
Hinton and LeCun say they would like
to end the dependence of today’s systems on explicit and extensive
training by people. Deep learning projects depend on an abundant supply
of data labeled to explain the task at hand—a major limitation in areas
such as medicine. Bengio highlights how, despite successes such as
better translation tools, the technology is not able to actually understand language.
None
of the trio claim to know how to solve those challenges. They advise
anyone hoping to make the next Turing-winning breakthrough in AI to
emulate their own willingness to ignore mainstream ideas. “They should
not follow the trend—which right now is deep learning,” Bengio says.
No comments:
Post a Comment