סמינר: Graduate Seminar
Developing Theoretical Frameworks for Understanding Continual Learning
| In continual learning, tasks are solved sequentially with the goal of maintaining high performance across the entire sequence. However, gradient-based algorithms often suffer from "catastrophic forgetting," where performance on earlier tasks deteriorates as new tasks are learned. While much of the existing research focuses on algorithmic solutions to mitigate forgetting, the theoretical understanding of these dynamics is less developed.
Our research aims to provide a theoretical foundation for understanding the dynamics of gradient algorithms in continual learning. We characterize the solutions obtained from continual learning and analyze various factors contributing to forgetting, including task similarity, task recurrence, overparameterization, and regularization. Specifically, we establish connections between continual learning with overparameterized linear models and the field of alternating projections, or more broadly, the Projection Onto Convex Sets (POCS) framework. Surprisingly, we find that learning additional tasks can actually alleviate forgetting. However, there exist adversarial cases where even after seeing an infinite number of tasks, forgetting can be arbitrarily bad. We also prove that when tasks recur, either cyclically or randomly, forgetting decreases over time and is universally upper bounded. Furthermore, we demonstrate that while overparameterization can help mitigate forgetting, it cannot fully eliminate it, and its impact on forgetting is highly dependent on task similarity. Our findings highlight significant differences between continual linear regression and classification, calling for further theoretical research specifically focused on classification settings. |
| His PhD has focused on theory of continual learning.
Ph.D. Under the supervision of Prof. Daniel Soudry.
|