Closed-loop Control for Online Continual Learning
Online class-incremental continual learning (CL) deals with the sequential task learning problem in a realistic non-stationary setting with a single-pass through of data. Replay-based CL methods have shown promising results in several online class-incremental continual learning benchmarks. However, these replay methods typically assume pre-defined and fixed replay dynamics, which is suboptimal. This paper introduces a closed-loop continual learning framework, which obtains a real-time feedback learning signal via an additional test memory and then adapts the replay dynamics accordingly. More specifically, we propose a reinforcement learning-based method to dynamically adjust replay hyperparameters online to balance the stability and plasticity trade-off in continual learning. To address the non-stationarity in the continual learning environment, we employ a Q function with task-specific and task-shared components to support fast adaptation. The proposed method is applied to improve state-of-the-art replay-based methods and achieves superior performance on popular benchmarks.
PDF Abstract
