Joelle Pineau, the Co-Managing Director of Facebook AI Research (FAIR) and head of FAIR Montreal, is one of six recipients of the Governor General’s Innovation Awards. The award recognizes Canadian leaders for their groundbreaking innovations and positive impact on the quality of life in the country. Pineau's research focuses on developing new models and algorithms for planning and learning in complex, partially observable domains. She is recognized for applying these algorithms to problems in robotics and in health care. Pineau is also a vocal advocate for increasing diversity among researchers and academics in the AI community.
In addition to leading Facebook AI’s research efforts in Montreal, she is also an associate professor at McGill University, where she co-directs the Reasoning and Learning Lab.
Her work at McGill has focused largely on applying AI to specific challenges in health care, whereas her Facebook AI projects concentrate on fundamental research and theoretical challenges in the field of AI. She has co-authored several papers in the last decade looking at ways to use AI to improve health care: Deep reinforcement learning that matters, Modeling glucagon action in patients with Type 1 diabetes, Learning robust features using deep learning for automatic seizure detection, Adaptive control of epileptiform excitability in an in vitro model of limbic seizures, Design and validation of an intelligent wheelchair towards a clinically-functional outcome, and Informing sequential clinical decision-making through reinforcement learning: an empirical study.
Pineau sat down to share her health care-related work and how AI can have a positive impact on the world.
I have had a long-term interest in applications of AI to health care. My Ph.D. thesis included work on building an assistant nursing robot. At the time, so many things were hard to accomplish efficiently, which today are much easier. Things like indoor robot navigation, reliable speech recognition, object recognition — these are all technologies that work today, which 15 years ago were nascent.
So when I joined McGill in 2004, I was looking for another opportunity to apply new AI technologies to health care. I started a project on smart wheelchairs with my McGill colleagues in engineering and in rehabilitation. Then, a short while after, there was an opportunity to look at the use of reinforcement learning to optimize neurostimulation strategies to reduce and prevent epileptic seizures, in collaboration with researchers in neuroscience.
In more recent years, we’ve started several new collaborations, many still ongoing, looking at various applications of machine learning in health care — including diabetes, cancer, heart disease, osteoporosis, and organ transplant. I’m grateful my position at Facebook AI still leaves me some time and a lot of freedom to continue these important projects. As a researcher, it’s always valuable to confront our theories with the real world, see how well our ideas really work, and see what assumptions we are making that might not hold up.
And at a more personal level, applications in health care seemed like a good way to eventually have a positive effect on people’s lives.
There are now many research labs and companies developing machine learning-based approaches for personalized medicine and adaptive treatment. Some of these I understand are based on work that we published. I think that, in some cases, the real impact on patients is still several years away. But it’s exciting to see that when, as a researcher, you present a proof of concept, there are other branches of the ecosystem that take over and find ways to apply, commercialize, and distribute the technology.
In many cases, there are problems well beyond the AI aspects that need to be addressed, such as energy efficiency, software reliability, and data privacy. In the case of our work in neurostimulation for epilepsy, the proof of concept was done in an in vitro model of epilepsy, which is much easier to control than a human brain.
Beyond the science, there are many other breakthroughs that are necessary. Some of them are on the ethical side, where we really need to start thinking more carefully about questions such as fairness, transparency, or privacy, and how we incorporate those considerations in the design and deployment of our AI systems, especially in a health care setting.
So many people! It is impossible to do this kind of cross-disciplinary work without several collaborators on the clinical side, and several passionate and brilliant students. Collaborators such as Massimo Avoli, Susan Murphy, Richard Gourdeau Joseph Cohen, and Robert Forget, to name just a few, were there in the early days when I started this program of research and were absolutely essential in conceiving and realizing the projects. And I could not have carried out these projects without the contributions of many very talented student researchers, as well.
It was very, very slow research! In some cases, we thought we had the AI models and algorithms ready to run the proof of concept, and then it still took two years to validate the model in the animal model of the disease. That said, what surprises me most is how quickly we went from a context where applications of AI in health care were quite esoteric to nowadays, where there is a huge amount of interest — so many clinical labs willing to share data, build collaborations, and see how we can make the most of the recent discoveries in AI. It’s exciting to see all the possibilities and the sheer variety of possible projects!
We have several new projects under way at McGill University, in collaboration with researchers in cancer and cardiology, both of which are new domains for me. So this requires becoming familiar with the de-identified data, the specific questions that doctors are trying to address, and carefully making a plan for an AI system that might make predictions or enhance decisions in a way that supports the clinical objectives. At Facebook AI, I’ve been working on more efficient and robust algorithms for reinforcement learning, and also on issues of reproducibility. Fundamental research like this is absolutely necessary for many of the projects we’re pursuing in health care. And separately, it’s been inspiring to see my Facebook AI colleagues working on fastMRI, a research project with the NYU School of Medicine that aims to speed up MRI scans by up to 10x.
Embrace small and slow research! It is essential to spend time understanding the data and the problem domain. And for the foreseeable future, there is really no shortcut for this. No amount of AI can tell you how to choose an outcome that is actually related to the patient’s well-being or how to integrate the predictions that the AI system is making into the clinical workflow.