COMPUTERS THAT THINK AND ACT LIKE HUMANS ARE THE GOAL FOR AI RESEARCHERS
Outside the Richards Building on campus one winter morning, it’s a few degrees above freezing. Inside, Prof. Graham Taylor reaches down to the heating control below the window to boost the temperature in his School of Engineering office. The device isn’t very smart, he says – and certainly not in comparison with the electronic climate controls at home. Installed in his century home in Guelph’s St. George’s Park neighbourhood, that Nest thermostat has learned his family’s daily and weekly routines and adjusts the temperature round-the-clock to suit. Maybe the house’s double-brick construction isn’t the best option for a frigid Canadian winter. But Taylor figures the computing smarts nestled in that 21st century thermostat help offset some of the built-in inconveniences of a 19th century dwelling. “We can have old-world charm, and we can make it more efficient by integrating machine learning,” he says.
That example brings home – in literal and figurative ways – Taylor’s studies since he arrived at U of G six years ago. Whether it’s your thermostat or other household fixtures and appliances, your phone or your (future) self-driving car, tomorrow’s smarter devices will increasingly rely on machine learning, a promising avenue being pursued by more artificial intelligence (AI) researchers in Canada and abroad all intent on making computers think and act more like humans. The U of G engineer hopes future breakthroughs – particularly in computer vision applications — will come from his own machine learning research group of some 20 students, post-docs and other investigators. Along the way, Taylor also aims to use so-called deep learning to help Canada further elevate its already-prominent international standing in the AI field.
From his office overlooking the Bullring and Reynolds Walk, he is tapped into regional and national networks of academics, companies and governments aiming to make this country a hub for artificial intelligence research and technology adoption. That goal is shared by Ottawa, which last year provided $125 million for a pan-Canadian AI strategy that encompasses research clusters in Toronto, Montreal and Edmonton. “Canada is really a leader worldwide when it comes to research in AI,” says Taylor.
Maintaining that lead by attracting and retaining AI talent and expertise is a key target of those networks, including the Vector Institute for Artificial Intelligence. Launched last year in Toronto, the institute aims to train the world’s largest cohort of graduate students in AI and to help commercialize research and technology. As its sole U of G member, Taylor brings his interests in machine learning and computer vision to that group. He’s also academic director of NextAI, a Toronto incubator for AI-enabled ventures in fields from finance and human resources to health care. There, he teaches machine learning and advises ventures in a program to launch businesses enabled by artificial intelligence. In 2016, the U of G professor was named as an Azrieli Global Scholar in the Learning in Machines and Brains program run by the Canadian Institute for Advanced Research (CIFAR) in Toronto. That program links researchers worldwide in efforts to create computers that can “think” as humans do – the basis of Taylor’s work in deep learning.
Taylor and other researchers hope to enable computers to teach themselves
How would you train a robot to tend your garden with all the unknowns from weather to soil conditions? Or how would you teach a self-driving car to navigate safely and efficiently through rush-hour traffic in Vancouver or Montreal, let alone Guelph? Following the conventional route of programming systems to account for every scenario and surprise, you’d quickly run up against roadblocks. “Humans can’t think through all the possibilities,” says Taylor. “You’d have to represent all these scenarios by programming, and include all sorts of rules and exceptions. To imagine all the possibilities is beyond our abilities.”
Instead, he and other researchers hope to enable computers to teach themselves. That challenge involves feeding in piles of data about a given scenario or application, and then allowing the machine’s interconnected neural networks to seek out patterns – a so-called “deep learning” process originally conceived to mimic the operations of neuronal networks in the human brain. That non-learning thermostat in his office understands “on” and “off” but not much more. By contrast, deep learning (what Taylor calls today’s “sexier” term for neural networks) involves more inputs, outputs and computation, as in a smart thermostat that uses those data to learn and adjust to the habits and patterns of building occupants. “You can build a system more efficiently by learning than by programming,” says Taylor, who credits his former PhD supervisors, Geoff Hinton and the late Sam Roweis, for much of this thinking.
Originally from London, Ont., Taylor worked with Hinton — considered a pioneer in the neural network approach to machine learning and now an adviser for both the Vector Institute and the CIFAR program – during grad studies in machine learning at the University of Toronto. After finishing his doctorate in 2009, he pursued a post-doc at New York University with Yann LeCun, co-director of the CIFAR program and now Facebook’s director of AI research; Taylor arrived in Guelph in 2012.
Along the way, he became fascinated by computer vision, or teaching computers to “see.” Working with his lab members at U of G and with various collaborators, including researchers in France, he hopes to train machines to recognize body pose and activity. Interpreting what others are doing is second nature for humans, but making sense of human poses and activities is a hugely complicated task for a computer.
Such research may find its way into various fields, including security and health care. In Taylor’s machine learning lab, master’s student Brendan Duke says his grandmother lost part of her vision after a recent fall. “She needs help to read the buttons on the microwave,” he says. “A lot of things would be more convenient if she had a better interface to the world.” His project partner, Alaa Ali, also a master’s student, says machine vision might be useful for airport surveillance. He says AI systems might also be used in human resources to help shortlist interview candidates and even to help conduct impartial interviews.
Post-doc Danesh Ramachandram works with Swift Medical, a Toronto startup company tracking wound healing for diabetic ulcers, surgical wounds or cancer lesions in hospitals or long-term care facilities. Treating and tending wounds costs Canada’s health-care system almost $4 billion a year; a machine-learning algorithm might better monitor healing and provide more accurate prognoses.
In an undergraduate project, Angus Galloway looked at using machines to help predict pending brain seizures. Now pursuing a master’s degree, he’s testing vulnerabilities in AI systems to protect them from hacking. Master’s student Terrance De Vries works on training algorithms to “know what they know” to improve machine learning, work that might apply in anything from medical diagnosis to driving navigation.
Reflecting Taylor’s proximity to plant and animal scientists on campus, the professor’s own research has ranged from use of aerial drones to assess soil health on farms, to sifting through video information to identify individual animals, to monitoring for a moth pest in orchards. Further afield, he’s lent his expertise in other ways. For an assignment with Google in California, he recently tested “continuous passive authentication” for keeping mobile phones secure. Instead of having to swipe your smartphone or enter a security code, wouldn’t it be more convenient if the device ID’d you by tracked your body language, movements, gait and other idiosyncrasies?
Referring to the range of projects in the machine learning research group, lab manager Brittany Reiche says, “There are endless possibilities of applications to help with everyday life.” Those everyday applications will multiply in our “Internet of Things” future, says Taylor. As machines mediate more of our lives at home, at work and in public places, they will converse among themselves in ways that will be largely invisible and inaudible to us. Tomorrow’s ever-smarter home thermostat will think for itself – but it will also talk to the other electronic brains around it. The smart fridge will keep tabs on its own contents and recommend suitable recipes, and maybe even maintain the shopping list and order the week’s groceries. Not that Taylor would mind: “I’d love to see deep learning reduce waste, enable us to eat healthier meals and expose us to new foods.”
He’s also eager to see how machines might help kids like his robot-crazy four-year-old learn STEM (science, technology, engineering and mathematics) topics. Recently he placed a deposit on a Tesla Model 3, an electric car with certain self-driving features. “I’m looking forward to a future where we don’t have to drive ourselves around. I think it will be safer.” He figures it will be at least another decade before self-driving cars become common. Whenever it happens, he says, he might be able to claim at least some credit by association. “It will be a deep learning car.”
Ensuring Data Security And Privacy In An AI World
Artificial intelligence – especially the automation of massive amounts of personal data needed for learning by AI systems — raises various security, privacy and ethical concerns, says computer science professor Rozita Dara.
Dara, who heads the data management and data governance research program at U of G, says AI may threaten information security. How to secure data against outright hacking as well as various uses – authorized or not – by governments, businesses and other organizations? She says artificial intelligence also raises privacy concerns. Might sensitive personal data be integrated or analyzed without users’ knowledge or consent? Might systems be smart enough to predict passwords or slip through security holes to gain access to data?
Paradoxically, the very systems that pose a threat may also prove the best tools for countering it, says Dara, whose interests range from big data analytics and data management, to information privacy, to ethical implications of technology. “We can use AI to process and analyze millions of data points to detect anomalies and suspicious behaviour online.” Her research might help ensure privacy through systems that understand users’ consent and preferences, and that share or protect data accordingly. She also studies the use of AI in data management and security technologies such as “smart contracts,” or intelligent software protocols that validate digital contracts.
AI can help ensure transparency and accountability in building and controlling these systems, says Dara, who was a privacy and information technology officer with the Office of the Information and Privacy Commissioner of Ontario before joining U of G. She calls for education in AI for researchers, practitioners and the public, as well as development of appropriate policies, procedures and government oversight.
As well as raising information security and privacy concerns, AI poses ethical issues, she says. How will AI affect human behaviour, values and social norms? How to account for potential biases in AI algorithms and the databases that systems learn from? Can we predict how intelligent systems will perform and prevent unintended consequences, including loss of human control?
“What concerns me are the massive amounts of data that are being collected in real time from billions of people around the world. No effective mechanism for governance, management and protections of data exist. We need to think about this as a society.”