Lethe: Plasticity-aware Active Forgetting for Resource-Efficient On-Device Continual Learning

On-device continual learning (CL) is becoming a critical paradigm for intelligent agents to adapt to evolving task streams in dynamic edge environments. However, existing CL methods primarily focus on optimizing learning performance, while neglecting the fundamental tension between ever-growing accumulated knowledge and limited model capacity on resource-constrained mobile devices, posing a significant challenge for model evolution on new task learning. To address this challenge, we reframe forgetting not as a problem but as a potential solution, and propose Lethe, a pLasticity-aware active forgetting scheme for on-device CL. Specifically, for fast model adaptation on new tasks, Lethe proposes forgetting-based plasticity injection to perform order-aware selective forgetting with neuro-inspired plasticity injection to accelerate model convergence. To prevent inefficient relearning of forgotten task re-encounter, Lethe introduces cost-aware forgetting recovery to facilitate low-cost knowledge recovery with deposited model parameters, and prevent frequent forgetting recovery based on task priority. Extensive evaluations demonstrate that Lethe consistently outperforms state-of-the-art solutions, improving model accuracy by up to 8.93% and reducing computation overhead by 53.63% on average.