SAN FRANCISCO – In 2004, Geoffrey Hinton doubled on his quest for a technical idea called a neural network.
It was a way for machines to see the world around them, recognize sound and understand natural language. But scientists had spent more than 50 years on the neural network concept, and machines couldn't do any of that.
Withdrawn by the Canadian Government, Dr. Hinton, a computer science professor at the University of Toronto, organized a new research team with several graduates who also tackled the concept. They included Yann LeCun, a professor at New York University and Yoshua Bengio at the University of Montreal.
Over the past ten years, the great idea that these researchers have reinvented how technology is being built has accelerated the development of facial recognition services, talking digital assistants, warehousing robots and self-driving cars . Dr. Hinton is now with Google, and Dr. LeCun works for Facebook. Dr. Bengio has inked agreements with IBM and Microsoft.
"What we've seen is nothing but a paradigm shift in science," said Oren Etzioni, CEO of the Allen Institute for Artificial Intelligence in Seattle and a prominent voice in the AI community. "History turned, and I'm in awe."
Loosely modeled on the network of neurons in the human brain, a neural network is a complex mathematical system that can learn discrete tasks by analyzing large amounts of data. For example, by analyzing thousands of old phone calls, it can learn to recognize spoken words.
This means that many techniques of artificial intelligence can be developed at a rate that was not possible before. Instead of coding behavior in systems by hand – a logical rule at a time – computer scientists can build technology that learns behavior to a large extent on their own.
The London-born Dr. Hinton, 71, first embraced the idea of a graduate student in the early 1970s, a time when most artificial intelligence scientists turned against it. Even his own Ph.D. adviser questioned the election.
"We met once a week," says Dr. Hinton in an interview. "Sometimes it ended in a screaming match, sometimes not."
Neural networks had a short resuscitation in the late 1980s and early 1990s. After one year of doctoral research with Dr. Hinton in Canada moved the Paris-born Dr. LeCun to AT & T's Bell Labs in New Jersey, where he designed a neural network that could read handwritten letters and numbers. An AT&T subsidiary sold the system to banks and at one time it read about 10 percent of all checks written in the US.
Although a neural network could read handwriting and help with any other tasks, it could not make much progress with great AI tasks such as recognizing faces and objects in photos, identifying spoken words and understanding the natural way one speaks.
"They worked well only when you had a lot of training data and there were few areas that had lots of training data," Dr. LeCun, 58, said.
But some researchers continued, including the Parisian-born Dr. Bengio, 55, who worked with Dr. LeCun at Bell Labs before joining the University of Montreal.
In 2004, Dr. Hinton a research program dedicated to what he called "neural calculation and adaptive perception". He offered Dr. Bengio and Dr. LeCun to join him.
By the end of the decade, the idea had gained its potential. In 2010, Dr. Hinton and his students Microsoft, IBM and Google on the limits of speech recognition. Then they did about the same with image recognition.
"He is a genius and knows how to create an impact after another," said Li Deng, a former speech researcher at Microsoft who took Dr. Hinton's ideas into the company. 19659002] Dr. Hinton's image recognition breakthrough was based on an algorithm developed by Dr. LeCun. At the end of 2013, Facebook employed N.Y.U. professor to build a research laboratory around the idea. Dr. Bengio opposed offers to join one of the big tech giants, but the research he supervised in Montreal helped drive the progress of systems designed to understand natural language and technology that can create false photos that can Distinguish from the real.
Although these systems unequivocally accelerated the development of artificial intelligence, they are still a very long way from true intelligence. But Drs. Hinton, LeCun and Bengio believe new ideas are coming.
"We need basic additions to this toolbox that we've created to reach machines that are on par with true human understanding," Dr. Bengio.