By: Krista Davidson
21 Jul, 2021
Irina Rish wants to solve one of AI’s biggest challenges
July 21, 2021
In the past several decades, there have been extraordinary advances in the way artificial intelligence (AI) machines learn. A major limitation, however, is in their ability to acquire new knowledge while retaining existing, relevant information — in a way that would be similar to how humans learn over the course of a lifetime. Canada CIFAR AI Chair Irina Rish is a pioneer in this emerging field known as continual lifelong learning.
Artificial intelligence (AI) systems consistently outperform and outpace humans in tasks such as identifying patterns and predicting outcomes. However, these systems often lack the versatility to apply the skills they acquire to new situations over time. Rish specializes in capturing the most robust and invariant properties across different data distributions to enable systems to continually learn over time, without further training.
“Systems are capable of adapting to new data, but when faced with new datasets and tasks, they often experience catastrophic forgetting, a process where new information washes out old information,” says Rish. “On the contrary, forgetting in human and animal brains seems to happen much more gradually, and less catastrophically.”
Rish is a faculty member at Mila, the Quebec AI Institute, and an associate professor in the Computer Science and Operations Department at the Université de Montréal. She was awarded a $34M Canada Excellence Research Chair in Autonomous AI in September 2020. She previously worked as a researcher at IBM Research in Yorktown Heights, NY.
She says there is a well-known “stability vs plasticity” trade-off in training AI systems based on neural networks, for example, a tradeoff between adapting and remembering. A system may sometimes retain information but be unable to adapt fast (too much stability, but not enough plasticity), while sometimes good adaptation may be unfortunately paired with quick forgetting. The ultimate goal of continual learning is to achieve fast adaptation while avoiding the catastrophic forgetting issue.
Continual lifelong learning systems are promising for a wide range of applications, particularly in AI for health, which is one of the focus areas for Rish. She is building models that use brain imaging data to identify different mental conditions, including disorders such as addiction and schizophrenia. Continual learning could allow us to avoid learning from scratch on new data by transferring knowledge learned from different patients and hospitals, and to ultimately extract robust brain activity patterns associated with mental conditions that are invariant across different datasets.
Continual lifelong learning also has applications for autonomous vehicles, which typically work really well in a controlled environment, but struggle to adapt in new environments and scenarios that it’s not familiar with.
“The ultimate goal is to develop robust algorithms that can learn on their own by reusing past knowledge, while continually adapting to novel tasks. However, at the same time, the algorithm has to be smart enough to detect situations when there is nothing similar or invariant between the old and new data, and expand the model if necessary,” says Rish. “Our research aims to bridge the gap between the generalization abilities of humans and AI systems.”
Irina Rish is a lecturer at the 2021 CIFAR Deep Learning + Reinforcement Learning Summer School.
Related Articles
-
What will be the impact of AI-assisted robotics on humanity?
February 15, 2023
-
CIFAR partners with Actua for Indigenous next-generation training in AI
November 08, 2022
-
CIFAR, Amii and the Vector Institute name eight new Canada CIFAR AI Chairs
October 27, 2022
-
DLRL 2022: That’s a wrap, and a return to in-person learning in 2023
September 28, 2022
Support Us
CIFAR is a registered charitable organization supported by the governments of Canada, Alberta and Quebec, as well as foundations, individuals, corporations and Canadian and international partner organizations.