Empowering machines with lifelong learning

As the Deep Learning + Reinforcement Learning Summer School (DLRL) kicks off its 16th annual edition, virtually, more than 300 students from across 45 countries will come together to cover the foundational knowledge and latest advancements in deep learning and reinforcement learning.

Many of the students in the program will go on to build careers in the field of machine learning and develop new products and services. Some will pioneer new breakthroughs and advances critical to our future.

Canada CIFAR AI Chair Sarath Chandar is a former student of the DLRL Summer School. Chandar is an academic core member of Mila, the Quebec Artificial Intelligence Institute, and an assistant professor at Polytechnique Montreal. He builds lifelong learning systems, which enable machines to learn beyond the initial data gathered in the algorithm they’re deployed in. His research has huge implications for any deployed machine learning systems, including digital personal-assistants, self-driving cars, and even medical applications, such as those that can detect cancer from medical imaging. 

A major limitation to current generations of machine learning systems is that once they are trained and deployed in a specific setting, they aren’t able to continue learning in the way that humans do.

“Some of these systems become outdated quickly and have to be retrained every few weeks. Every time you have to retrain a system, it uses a lot of computational resources, time and data,” explains Chandar.

Empowering systems with lifelong learning could have life-saving consequences. For example, it could help self-driving cars adapt to new scenarios and situations, for example, driving in inclement weather conditions or new routes it may not be familiar with. For medical imaging, it could be the detection of abnormalities it hasn’t yet been trained on.

“Lifelong learning is one of the ultimate promises of AI and there are several fundamental issues that need to be addressed before we build an ideal lifelong learning system.”, says Chandar. His lab works on both the fundamental research and the applications of lifelong learning in computer vision and natural language processing.

Prior to becoming a faculty chair and an acclaimed researcher in machine learning, Chandar attended two summer schools - the first in Montréal as a PhD student at Université de Montréal, where he was advised by deep learning pioneers Yoshua Bengio and Hugo Larochelle. Both Bengio and Larochelle are CIFAR Fellows of the Learning in Machines & Brains Program and Canada CIFAR AI Chairs.

In 2019, he attended the second school in Edmonton after he had just accepted his first faculty position. He says he met a lot of future colleagues and has recruited students from the school to join his research team.

“When I attended both DLRL Summer Schools, I didn’t miss a single lecture,” he says. “For me, the appeal is that you get to meet people with different experiences and see how they are applying machine learning to their research, but also you get a crash course on a topic you may know a little or a lot about. It’s important to get out of your own research area and see what advances are being made in other areas in machine learning.”

Chandar adds that for students coming from outside of Canada, it’s a chance to witness first-hand Canada’s world-class academic institutions and research labs. 

“I can’t think of another country in the world that collaborates the way Canada does,” he says. “You see students and researchers from across multiple universities collaborating on ideas. You don’t see that anywhere else.”