Professor Pascal Poupart got hooked on computer science early. It began decades ago when he took a programming course and fell in love with the discipline. But perhaps even he didn’t foresee how this nascent interest in coding would lead to developing interactive artificial intelligence that improves hockey and physical and mental health.
“Artificial intelligence or AI is an area of research where we try to get a computer to do what’s best in achieving a goal,” said Poupart, a leading researcher in machine learning and reasoning under uncertainty, health informatics and natural language understanding. “One way to think about it is, computers will take some inputs — voice, text, images and so on — then it’ll produce some outputs, which could be voice, text, images or in the case of robotic systems perhaps a controlled movement.”
When thinking about AI it’s helpful to picture a computer as though it were an agent, like a human being. With AI, the computer scientist works to optimize outputs so the computer will achieve the best outcome across a range of circumstances. In the early days, the field was inspired by how humans did things — how humans think, how humans act. Today, however, computer scientists working on AI consider the best way of acting, behaving or thinking.
“We treat humans as an example of a system that works but one that might not necessarily be optimal. That’s why in many circumstances a computer can outperform a human because they are not trying to simply imitate human behaviour. Computers are trying to go further.”
Poupart is putting this to the test through a collaboration with HockeyTech, a local company that analyzes data from hockey games.
“With data automation all kinds of analyses are possible for coaches, players and fans to better understand how play can be improved.”
Automating recognition of events is an interesting problem because hockey is a quick-paced game and it’s not easy to know with certainty what’s happening at any given point — for example, who’s in possession of the puck, when a pass is successful, or even to recognize shots on goal.
“We’re using transmitters — one in the puck and one on each player so we know their positions — and we’re looking into using a camera in the ceiling of the rink to give a top-down view to complement the positional data,” he said.
“In professional hockey, scorers watch the game and record the shots. We’re doing this automatically. Once automated the cost of the system will go down and it could be deployed everywhere — not just in professional arenas but also at amateur rinks.”
This AI system should improve the quality of hockey but it could also help talent scouts to determine which players to recruit and help coaches to better understand what makes one player superior to another.
In another project, Professor Poupart is working with researchers in Waterloo’s Department of Kinesiology to help people stay physically active. Lots of people have smart phones and wearable devices that will remind them to walk a certain number of steps or to break up sedentary periods during the day.
“When wearable devices like fitness and smart watches became available everybody was rushing to buy one,” he said. “People typically wear one for a few weeks, but after the novelty is gone they may lose interest.”
Poupart wants to use AI to inspire a change in behaviour.
“We’re social beings and we tend to be more motivated when others are involved,” he said. “If you’re going to coach someone else, presumably you’re going to be more conscious of what needs to be done and you might be more effective in achieving your own goals. We are carrying out an experiment to see whether this makes a difference and hopefully lead to a sustainable change in behaviour.”
By designing an AI system that experiments with different approaches to guide and coach people, it could learn over time what works to refine that guidance.
“There are some techniques from a machine-learning perspective that we can explore to achieve that,” he says. “This leads toward a broader aspect of my research, what I call interactive machine learning. Machine learning helps extract useful information from data, but for many kinds of scenarios — when a machine is supposed to learn about a person, his or her preferences and habits — we don’t typically have lots of data to work with.”
People expect a machine to learn or to understand them after a few trials, so much effort is being placed on developing interactive machine-learning techniques that don’t need large quantities of data by taking into account that a human is in the loop. Machines can be trained to do something provided that we give them labelled data, essentially telling the machine what was the right thing to do. When it makes predictions or a mistake the user can provide feedback, which the machine then uses to refine its behaviour.
“If a person asks a question and the machine responds but doesn’t provide a good answer or is off topic then it’s likely the person will say something like, "Hey, you didn’t answer my question or this is not what I was asking," he explains. “You give implicit feedback and now the question becomes how can we do machine learning that takes that into account.”
Aspects of this research direction are being pursued with colleague Igor Grossman in Waterloo’s Department of Psychology to recognize levels of stress based on physiological measures such as heart rate and skin sweatiness that are inferred from a wearable sensor.
“We’d like to understand the triggers of stress among students, one of the most important issues they face. We are looking at machine-learning techniques that map data from a wearable sensor to assessments that initially would be provided by a user to indicate that the person currently is or is not experiencing stress.”
Professor Poupart is optimistic the system could also be used in the workplace to help reduce stress. “Some people end up on long-term disability benefits because of mental illness,” he explains. “Mental illness is a broad thing that isn’t well understood, but a good component of it could be related to stress. If we better understand triggers of stress we should be able to reduce it.”