Continual Learning Working Group: Lea Duncker

Title: Task-dependent low-dimensional population dynamics for robustness and learning

Abstract: Biological systems face dynamic environments that require flexibly deploying learned skills and continual learning of new tasks. It is not well understood how these systems balance the tension between flexibility for learning and robustness for memory of previous behaviors. Neural activity underlying single, highly controlled experimental tasks has repeatedly been observed to exhibit low-dimensional structure. However, it is unclear how this organization arises and is maintained throughout learning, and how it might differ when networks are exposed to multiple tasks. In this talk, I will present work on a continual learning rule designed to minimize interference between sequentially learned tasks in recurrent networks. The learning rule preserves network dynamics within activity-defined low-dimensional subspaces used for previously learned tasks. It encourages recurrent dynamics associated with interfering tasks to explore orthogonal subspaces. Employing a set of tasks used in neuroscience, I will show that this approach can successfully eliminate catastrophic interference, while allowing for reuse of similar low-dimensional dynamics across similar tasks. This possibility for shared computation allows for faster learning during sequential training. Finally, I will highlight limitations of this approach in fully exploiting task-similarity for optimal re-use of previously learned solutions, and outline new work we are starting in my group now to address this.

Zoom Link: https://columbiauniversity.zoom.us/j/97176853843?pwd=VLZdh6yqHBcOQhdf816lkN5ByIpIsF.1