Grant: Canada Research Chairs

A key aspect of human intelligence is the ability to quickly adapt to solving new problems in a constantly evolving world. Humans do this by accumulating knowledge over their lifetimes, and composing and reusing this knowledge to perform new tasks. Despite significant progress in machine learning research, powered by advances in deep learning, adapting to changes in the world continues to be an important limitation. This research program proposes addressing this limitation by exploring the paradigm of lifelong learning. Lifelong learning systems learn multiple tasks continually over time and should have the ability to transfer previously gained knowledge to effectively and efficiently learn new tasks. Developing effective lifelong learning systems can have a significant impact on theoretical and practical aspects of machine learning. To this end, this research program will develop novel model architectures that leverage modularity in design and long-term memory to grow their knowledge over time. The program also proposes working on developing new training tools such as optimizers designed for lifelong learning, ways to learn new tasks from natural language instruction and interaction, and ways to leverage recent advances in pre-training toward building systems that can be continually adapted to new tasks. These complementary research directions of developing new architectures and new training tools can help overcome fundamental limitations of modern neural network-based learning systems such as catastrophic forgetting and capacity saturation, and enable a system to transfer the knowledge gained from learning a new task to improve on tasks it has seen before, and tasks it will see in the future. From a practical standpoint, developing lifelong learning systems can help any machine learning system stay up to date and adapt to changes in the world without having to undergo the expensive process of periodic retraining on newly collected data. Advances in lifelong learning systems can also enable practical applications such as dialogue systems that can become more personalized over the course of interaction with a user, and drug discovery systems that can accumulate knowledge over time and get better at finding drugs for new targets. Developing ways to continually adapt learning systems can also make large-scale pre-trained models more accessible and resource-efficient, and have a significant impact in any setting in which they are deployed.