Continual Learning at the Edge
We study how ML systems deployed on resource-constrained, distributed devices can adapt over time without storing raw past data and without overwriting previously acquired knowledge. The project develops a unified replay-free, projection-based framework for three core deployment challenges: efficiency, through compact and quantized task subspaces for low-SWaP platforms; controllability, through architectural and activation-level disentanglement that enables safe removal of obsolete task knowledge; and collaboration, through federated continual learning across heterogeneous clients without raw-data sharing or task labels at inference.
