Grant: Discovery Launch Supplement

Autonomous systems are increasingly integrated into daily life, assisting in our homes and vehicles. These systems are designed to work for and alongside human users, necessitating personalization to individual preferences and unique environments. Albeit the prevailing approach to developing autonomous systems, pre-training with general knowledge proves impractical for personalization, as it is impossible to accurately predict the diverse set of users' specific needs and environments. Another challenge for pre-designed systems is long-term deployment, as users' preferences and environments can change over time. The proposed research program addresses this by developing continual learning (CL) methods for autonomous systems to continually learn user preferences and environments through interactions with their users. A key problem for CL models is catastrophic forgetting (CF), where the model forgets previously learned knowledge when learning new information. Most CL models for object classification that mitigate CF require a large number of training images per object class which is impractical for users when teaching autonomous systems. Orthogonal to this, active learning (AL) models select only a small subset of training samples to be labeled, reducing the required user effort. Accordingly, this research program focuses on developing continual active learning (CAL) methods that can continually 1) learn the user-provided information in a computationally efficient way and mitigate CF 2) determine the most important information to perform a task, 3) determine when to ask users for information based on user preferences. My PhD research took the first step in this direction and developed a novel CAL method, that combined AL and CL methods to continually determine the most uncertain samples and mitigate CF. Despite achieving state-of-the-art performance, this method did not address all three challenges, as it is computationally expensive and uses fixed AL methods that do not adapt to users' preferences. This research program will address all three challenges by developing adaptive AL methods using Bayesian neural networks that adapt to learn objects that best help the system adhere to users' needs. To mitigate forgetting, computationally efficient methods for updating a select portion of network weights will be developed. Reinforcement learning policies will be trained to determine the most appropriate time to query users for information. In addition to testing CAL methods on object classification datasets, we will perform long-term human-computer interaction studies to understand the performance of CAL methods when directly learning from human users, and users' perceptions of these methods. This research program will fill critical gaps in the field of personalized autonomous systems by developing CAL methods, and train HQP to be equipped to contribute to the fast-growing Canadian autonomous system sectors, including robotics and autonomous vehicles.