Marzyeh Ghassemi is a Canada CIFAR AI Chair and a faculty member at the Vector Institute. She was recruited from the U.S. to Canada and the University of Toronto, where she is a professor in the departments of computer science and medicine.

Chris Pal is a Canada CIFAR AI Chair and a faculty member at Mila. He is an associate professor at Polytechnique Montréal, and an adjunct professor with the department of computer science and operations research at Université de Montréal and a principal research scientist at Element AI.

Martha White is a Canada CIFAR AI Chair and a faculty member with Amii. White was recruited from the U.S. to the University of Alberta, where she is an assistant professor in the department of computing science. She is also the director of the University of Alberta Reinforcement Learning and Artificial Intelligence Lab.

Reach: Tell us about your area of research?

Martha White: I work on reinforcement learning, which is a form of machine learning. The idea is that a [software] agent doesn’t get to know much information about its world and it attempts to learn how to take action in that world through trial and error interaction. My particular expertise in this area is looking at how the agent can build up a representation or knowledge about its environment so that it can learn how it should take action in a more intelligent way.

Marzyeh Ghassemi: I am focused on identifying what kind of models are healthy in a health care setting. We’ve seen great advances in application areas like vision, speech and natural language but we’re still looking for structures and models that work well based on how health data is generated. I am trying to harness some of the power of the observational data that we have in the Canadian health care system and use machine learning and causal inference methods to create more accurate and scalable recommendations for health.

Chris Pal: I’ve been doing AI research for over 20 years. I’ve spent a lot of time on computer vision which has applications for medical imaging. I’ve conducted research on segmentation, specifically segmenting brain tumours in medical images. We’re looking at deep learning to improve segmentation accuracy of tissue types, which is very useful for clinical decision making.

Investing in training is one of the smartest moves that any country can make." — Marzyeh Ghassemi

Marzyeh Ghassemi: Right now we are moving from a space in which everybody is amazed at what AI can do — whether it’s speech recognition, predicting mortality or self-driving cars — to a space in which we’re starting to care about the nitty-gritty details of how possible technologies might get implemented. That’s not something that incentivizes all researchers. AI and machine learning models can work in many settings. We can understand whether machine learning can make predictions about people of all races with equal accuracy, and whether it can be deployed in different countries that have different health care practices without impacting underlying health care.

Martha White: We will start to see that machine learning and reinforcement learning become part of the solution strategies people use to tackle regular problems on a daily basis. Similar to how kids in school today are taking computer programming early on and computer science is seen as a standard skill that many people have, we’ll start thinking of modelling and prediction tools as a standard skill set, and we should absolutely be introducing it earlier at the undergraduate level.

Chris Pal: There’s tremendous potential for AI in health. There are a number of organizational problems that can benefit from AI if we get it right, particularly in medicine. How do we get data out of a clinical setting into the hands of researchers in a way that respects privacy and enables biases to be corrected? The Canada CIFAR AI Chair program helps because researchers can focus on fundamental research and also have access to many resources critical to fully realizing the potential of AI.

Marzyeh Ghassemi: When I was at NeurIPS [the Conference on Neural Information Processing Systems], there was a set of playing cards mimicking Cards Against Humanity and one card asked: “Have you collaborated with the Canadian mafia?” It was a tongue-in-cheek reference, but it pointed out that many of the advances led by researchers in Canada have paid off in terms of neural network architectures and inference methods. This is a result of people being able to work on ideas that were not hot at the time. The payoff that we’re experiencing now is the result of investing early on in allowing people to do the research that really speaks to them.

Chris Pal: It’s a really critical time period where a lot of experienced researchers would have moved on to industrial research activities, resulting in a lot of attrition in academia. Because there is a lot of support in Canada around academic research in AI, more people are working on academic research, and the timing of that is critical. It means that more students will be trained in Canada.

In our case with the creation of the AI Institutes, Amii, Mila and Vector — that’s already starting to have a very positive effect on the environment under which many people with related, complementary interests are co-located. That allows for a lot of the critical mass effects that are not possible in smaller groups. I have collaborations with people outside of Mila’s traditional areas because of the critical mass of talent we have in Canada. We’re on the radar globally. Today, everyone is interested in talking about what AI might be able to do and it’s encouraging people to think innovatively, and that’s a very positive thing.” — Chris Pal

Martha White: When AI came on the horizon, fundamental research continued to be funded reasonably well in Canada. And that had a compounding effect because more researchers wanted to come here for research.

Reach: The CIFAR Pan-Canadian AI Strategy focuses on supporting talent and training for the next generation, whereas other national strategies support commercialization and take a more industrial focus. Is ours a good approach?

Marzyeh Ghassemi: Investing in training is one of the smartest moves that any country can make. The kinds of people that are going to contribute most to fundamental questions and research want to work on challenging problems and they want to know what other like-minded people have done. It’s a huge draw. Bringing in smart people allows them to innovate. That’s going to pay off disproportionately more than any other investment you could make.

Martha White: I agree completely. Coming from the U.S. system, where I was for a couple of years, I really feel like the focus in Canada is much more on highly qualified personnel, training individuals and funding people rather than projects. In other countries, there is a lot less funding for students directly and that calls for more specific research projects that tend to pigeonhole researchers, rather than permitting an open-ended training environment where people can grow and develop in a lot of different ways.

Chris Pal: To fuel startups you need people with a lot of skills, which you almost only get by doing. A lot of times you learn those skills through graduate research or by working in advanced industrial research labs. When students go into a strong graduate program, that’s when they start to understand what innovation is. You want to invest in people rather than innovation. That’s a powerful component of our funding structure in Canada.

Reach: What do your next five years as a Canada CIFAR AI Chair look like?

Marzyeh Ghassemi: In the next five years, I would like to see Toronto, and Canada, really start to lead the charge in machine learning for health. We have some amazing raw resources. We have an inclusive health care system that all people have access to; that’s not true of other countries. We also have fantastic hospitals in the Greater Toronto area with access to data from diverse communities and globally ranked medical schools. One of the things I’m most excited to do is to leverage my AI Chair position towards bridging those communities. Those who train in the medical field generally tend to publish in separate venues and are not always well connected to training and opportunities for publishing research with computer scientists.

I was on the fence when choosing which university to go to, and what really convinced me to come to the University of Toronto and Vector was the opportunity to work with Roger Grosse and Jimmy Ba [also Canada CIFAR AI Chairs]. The fact that I was going to work with colleagues of that calibre and could advise and brainstorm with students created a strong pull. All of the graduate students I’m admitting [into my research projects] have been offered positions at MIT, Stanford University or Carnegie Mellon University. They are choosing us [University of Toronto] over some very competitive schools because of the access to the talent they will have and the ability to work within a larger research mandate.

Chris Pal: The notion of how CIFAR funding can help a professor connect the dots to applications is a concrete impact. As Canada CIFAR AI Chairs, we have funding that is very flexible and stable for five years. It allows us to bring on a PhD student and focus attention on research that will have an impact on the real world. The AI Chairs program allows researchers to maximize impact in whatever way is the most appropriate.

The focus in Canada is much more on highly qualified personnel, training individuals and funding people rather than projects." — Martha White

Martha White: Stable funding means that you can work on problems that are important to you. Over the next five years, I’m hoping to make connections that will help put reinforcement learning in practice for more problems. There are a few fundamental challenges but I think we are on the cusp of getting past some of those challenges where we can deploy in more settings. I am investing in this physical system and we’ll be trying different algorithms on that system. I’m hoping within the next five years to get some insight out into the real world with some robust algorithms.

Reach: AI is quite the buzzword these days. Should we be excited about AI or is it just hype?

Marzyeh Ghassemi: Yes, we should be excited. As a community we need to make sure we mitigate some of the damages from that hype. We, as researchers, have to be very realistic about the challenges and the possible negative impact because all new tools have benefits and drawbacks. And we can’t account for certain information without bringing humans into the equation. For example, medical doctors with a decade of training need to be involved in assessing the AI tools used in medical settings. We also need to remember that people are biased and therefore models will also be biased.

Martha White: I’m excited that our society is moving to a data-driven one where we actually recognize the utility of gathering data. But it puts the onus on the machine learning community to avoid over-claiming too much because that’s never been a good idea.

Chris Pal: I think hype can have a positive impact to the extent that it gets people excited about innovating. It helps people open their mind to doing things in a different ways. One positive aspect of the AI hype is that even a decade ago, if you wanted to get in front of industrial partners, it was almost impossible. Today, everyone is interested in talking about what AI might be able to do. It’s encouraging people to think innovatively, and that’s a very positive thing.