Experimental Evaluation of the REM-Sleep Protocol for Continual Learning

This study addresses the pervasive challenge of Catastrophic Forgetting in Continual Learning (CL) by introducing a biologically-inspired "REM-Sleep" protocol. Unlike standard online rehearsal methods such as Experience Replay (ER), this proposed architecture decouples the learning process into distinct Awake (acquisition) and Sleep (consolidation) phases. By utilizing noisy replay and synaptic homeostasis during the offline sleep phase, the protocol aims to stabilize memory representations without interfering with ongoing learning. Experimental evaluations on a 5-task Split-Digits benchmark demonstrate that the REM-Sleep protocol significantly outperforms standard baselines in memory stability, achieving a 9% relative reduction in Average Forgetting compared to ER (F=0.171 vs. 0.188). Furthermore, the research includes a critical ablation study on sleep duration, identifying an optimal consolidation window of 10 epochs while demonstrating the detrimental effects of both insufficient and excessive sleep. These findings present a compelling case for offline consolidation strategies as a viable path toward robust artificial neural networks capable of lifelong learning.