Turing Award recipient and McGill alumnus Yoshua Bengio, best known for his seminal work on artificial intelligence (AI), is the latest recipient of the Killam Prize, a $100,000 award given to outstanding scholars. Bengio is the founder of the Montreal Institute for Learning Algorithms (Mila), a world-renowned AI research institute.
He attributes his success, in part, to Canada’s investment in curiosity-driven research.
“We have been successful in AI […] in great part because we have a research funding system that allows professors to learn some subject that no one cares about, but that may have great potential in the future,” Bengio said in an interview with The McGill Tribune. “I would like there to be more of it, but at the same time, I realize that we’re really lucky in Canada to have the system we have.”
Bengio has been a vocal opponent of autonomous, AI-driven weapons. In 2017, he co-authored an open letter with other prominent AI researchers calling for the Canadian government to support an international ban on the use of lethal autonomous weapons—which operate without meaningful human control—at the United Nations.
Another recent concern with AI is deepfakes, which are artificial—but highly realistic—videos created using generative adversarial networks (GANs). GANs were developed by Bengio and colleagues in 2014, and essentially use two competing neural networks to produce novel data that resembles existing data. In light of a recent fake video of Nancy Pelosi, digital policy pundits fear that both crudely manipulated videos and more sophisticated deepfakes will be used in further smear campaigns against political candidates.
“It’s an issue right now in the courts and media—people take for granted that if there’s a picture, it’s true,” Bengio said.
According to Bengio, deepfakes are currently easy to detect, but the arms race between deepfake producers and detection software will continue. A researcher behind a recently developed method of identifying deepfakes conceded that although it will be especially difficult, deepfake-makers will likely adapt to the new technique.
Deepfake texts also pose a problem. The creators of a revolutionary AI text-generator called the GPT-2 model refused to release their entire research for fear of potential misuse. Despite these concerns, though, Bengio remains optimistic about the future of AI.
“Think about how access to the Internet is increasing our ability to have access to more knowledge,” Bengio said. “It is going to depend on how we choose to deploy these [tools]. If we do it in a mindful way rather than just driven by maximizing profits, I think we could do something pretty good for society.”
Recent strides made in AI research are advancing various areas of science and technology. Last year, a study published in Nature Medicine revealed that Google DeepMind’s AI can interpret 3-D eye scans and make more efficient referral decisions than expert physicians. DeepMind is also being trained on weather forecast data to predict the power output of wind turbines up to 36 hours in advance, leading to increased energy efficiency.
Bengio believes that the future of AI will benefit from collaboration.
“Having more people at the intersection of some domain and AI is something that society will need a lot in the future,” Bengio said.
He also promoted the power of self study, urging university students to leverage free and open-source tools for learning instead of simply relying on professors or courses. Bengio offered a final piece of advice for students interested in pursuing a research-oriented career.
“I hope people will take the time to think because big advances in research actually come from a place where you have quiet and time to think and you’re not in the rush of producing something tomorrow,” he said.