DeepMind goes to Alberta for first international lab
(Bloomberg) -- DeepMind, the London-based artificial intelligence company, is hiring three prominent computer scientists from the University of Alberta in Edmonton, Canada, to establish its first research facility outside the U.K.
The new lab will be headed by Rich Sutton, a leading expert in reinforcement learning, a form of machine learning in which software learns by trial and error to maximize a reward. The company is also hiring Michael Bowling, a professor who has used reinforcement learning to train software capable of playing poker better than many of the world’s top professionals, and Patrick Pilarski, who has studied the creation of AI-enabled artificial limbs.
As the race to develop ever more advanced machine learning capabilities accelerates, wealthy tech companies from Facebook to Uber have increasingly raided university computer science departments to gain an edge on rivals, a trend that troubles some academics who fear the brain drain is making it harder to train the next generation of researchers.
DeepMind is owned by Alphabet Inc., the parent company of Google. It is best known for developing AlphaGo, software that has beaten the world’s best players at the strategy game Go, an achievement considered a major milestone in computer science.
The announcement of the Edmonton lab comes two days after the U.K.’s top privacy watchdog found that a London hospital that partnered with DeepMind to develop a mobile app had broken the law by giving the company access to 1.6 million medical records without the patients’ consent.
Sutton has a long relationship with DeepMind. He met Demis Hassabis, DeepMind’s co-founder and chief executive officer, and his co-founder Shane Legg, at a conference in Geneva in 2010 before they started the company. Sutton has also served as DeepMind's first scientific advisor and helped to supervise DeepMind researcher David Silver, who went on to lead the AlphaGo development team, when Silver was studying for his doctorate at the University of Alberta. Two other members of the AlphaGo team --Aja Huang and Marc Lanctot -- also studied at the University of Alberta.
About a dozen of University of Alberta's alumni currently work at DeepMind, Hassabis said in an interview with Bloomberg. Bowling built a platform that allowed researchers to develop software that would play old Atari computer games, work DeepMind built upon in 2015 when it created AI capable of rapidly learning, with no prior knowledge, to play many Atari games far better than any human can.
Hassabis and Sutton said it was too early to say exactly what projects the new team would pursue. The three University of Alberta researchers will initially be joined in DeepMind's new lab by seven other researchers, including some who are relocating from the U.S. and the Czech Republic. Hassabis said the company is currently scouting for office space in Edmonton close to the university.
In addition to working for DeepMind's new lab, the three researchers will continue to teach at the university and to supervise graduate students, a statement from the University of Alberta said. Sutton said they would be paid as part-time professors by the university as well as drawing salaries from DeepMind.
While technology companies have always hired recent graduates in computer science to staff their own research and development departments, in the past five years these businesses have built large artificial intelligence research teams by luring away top professors with hefty pay packages.
Google tapped the University of Toronto's Geoff Hinton, who pioneered the use of neural networks, a kind of artificial intelligence that mimics how part of the brain works, while Facebook hired Yann LeCun, a student of Hinton's who was a professor at New York University. Chinese web giant Baidu hired Andrew Ng, a top AI academic from Stanford University, in 2014, although Ng resigned earlier this year to work on his own AI project. In the most extreme case, in 2015, Uber hired about 50 researchers from Carnegie Mellon University's robotics lab in Pittsburgh-- about a third of the center's entire staff -- to work on the company's self-driving car efforts. In May this year, Uber snapped up Raquel Urtasun, an associate professor from the University of Toronto considered a rising star in the field.
The trend alarms some academics, who worry corporate priorities won't always align with important basic research topics and that the field is in danger of eating its own feedstock. "This is creating a concern that the huge demand for supervision of graduate students won't be met, which in the long run is bad for industry too," Yoshua Bengio, an artificial intelligence researcher at the University of Montreal, said in an e-mail.
Bengio, along with Hinton and LeCun, is considered one of the godfathers of the use of neural networks. He said top universities were having to turn away promising Ph.D. students because they did not have enough professors to supervise them. He added that the best graduate students were lured away by tech companies before they could complete their studies. "If in the future these industries stop doing as much basic research as they do now (this has happened in the past), these people will be stuck in industry," he said. "This phenomenon could also be detrimental to academia and basic research in the long run."
Bengio said he was more supportive of arrangements, like the one DeepMind has struck with the three University of Alberta professors, where the professors maintain close ties with a university and continue to teach and supervise graduate students. "I believe those companies have a duty, when they strike such deals, to find ways to give back to academia, to help recruit other (maybe younger) professors and fund more graduate students," he said.
Jonathan Schaeffer, dean of the faculty of science at the University of Alberta and also an artificial intelligence researcher, said rather than hurting the computer science department, it would help the university hire faculty and attract promising students. "You have to understand, this is Edmonton, Alberta, Canada -- we are not on the beaten path," he said. Schaeffer estimated that fewer than five of the 75 graduate students who have trained under him in Edmonton have remained in the city. DeepMind's lab, he said, could change that.