Continual Learning

Continual learning addresses the fundamental challenge of training machine learning models on sequential, non-stationary data streams while preserving knowledge from previous tasks. My research in this area focuses on:

Theoretical Foundations: Formalizing the stability-plasticity trade-off that governs how models balance retaining old knowledge versus acquiring new capabilities. Our NeurIPS 2021 work established mathematical bounds on the generalization-forgetting trade-off, providing principled guidance for algorithm design.

Dynamic Programming Approaches: Developing algorithms that treat continual learning as an optimal control problem, enabling principled decision-making about when to update, consolidate, or protect learned representations. This perspective connects classical control theory with modern deep learning.

Scientific Applications: Applying continual learning to real-world scientific problems including defect identification in materials science (coherent diffraction imaging at synchrotron facilities) and chemical reaction yield prediction using large language models.

Key contributions include novel regularization strategies, graph-based continual learning methods for dynamic data structures, and uncertainty-aware approaches for detecting when models encounter distribution shifts.


Publications