Entropy-Gated Plasticity Regulation for Continual Learning Without Episodic Memory

Catastrophic forgetting remains a central challenge in continual learning. Existing approaches based on regularization or episodic memory offer effective solutions but introduce specific computational and memory tradeoffs. This work presents Entropy-Gated Plasticity Regulation (EGPR), a lightweight mechanism that modulates learning rates based on the temporal derivative of predictive entropy. By calibrating the entropy distribution on an initial task and gating subsequent weight updates through a sigmoid function of the entropy z-score, EGPR selectively inhibits plasticity when the network encounters out-of-distribution data. A complementary mechanism, Depth-Modulated Plasticity (DMP), applies layer-wise scaling to preserve feature representations in early layers while permitting adaptation in later layers. Empirical evaluation across six benchmarks (Split-MNIST, Permuted-MNIST, Rotated-MNIST, Split-FashionMNIST, Split-CIFAR10, and Split-CIFAR100) with five random seeds demonstrates that EGPR achieves significantly higher retention than Elastic Weight Consolidation (EWC) in five of six benchmarks (one-sided Wilcoxon signed-rank test, p = 0.031), without requiring episodic memory or explicit task boundaries. On Split-MNIST, EGPR achieves 85.7% retention compared to 10.0% for EWC, without any episodic memory. Furthermore, combining EGPR with experience replay yields statistically significant retention improvements over standard replay alone in all six benchmarks (p = 0.031), with gains ranging from +3.2 to +45.3 percentage points. On Split-CIFAR10, EGPR+Replay achieves 59.5% retention compared to 14.2% for replay alone, establishing EGPR as a highly effective complementary module for existing replay systems. The method's boundary conditions, notably the Semantic Distance Deadlock on complex visual domains, are documented and analyzed to inform future work.