Please register for the seminar here.

Abstract: 

Modern machine learning systems must navigate the dual demands of forgetting and retaining knowledge—whether to comply with data deletion requests or to adapt continually in dynamic environments. In this talk, I will present two of our recent works that address these challenges using principles from game theory and geometry.

The first part of the talk focuses on machine unlearning, where the goal is to selectively erase behaviors without retraining from scratch. We formulate the problem as a cooperative two-player game between a forgetting player and a preservation player, each proposing gradients toward their respective objectives. Drawing on Nash bargaining theory, we derive a closed-form update rule that guides the model toward a Pareto-optimal solution. We demonstrate the effectiveness of the resulting algorithm across both classification and generative tasks.

In the second part, I turn to the challenge of lifelong learning in State-Space Models (SSMs), particularly in the exemplar-free setting. We introduce a geometry-driven regularization method that operates on the infinite-dimensional Grassmannian. This framework captures the long-term dynamical behavior of SSMs over an infinite horizon and uses it to preserve prior knowledge—without relying on memory buffers or data replay. Our method achieves strong performance across multiple vision benchmarks, significantly reducing catastrophic forgetting while maintaining adaptability to new tasks.