MPCS: Neuroplastic Continual Learning via Multi-Component Plasticity and Topology-Aware EWC

Continual learning systems struggle to learn new tasks without forgetting old ones. A new paper introduces MPCS (Multi-Plasticity Continual System), an architecture that integrates 11 complementary mechanisms to balance plasticity (learning new knowledge) and stability (retaining old knowledge). These include task-driven neurogenesis, Fourier-encoded inputs, EWC regularization, meta-replay, mixed consolidation, hybrid gating, synapse pruning/regeneration, Hebbian updates, task similarity routing, adaptive growth control, and continuous neuron importance tracking.

Evaluated on MEP-BENCH—a 31-task benchmark spanning regression, classification, logic, and mixed domains—MPCS achieved a Normalized Efficiency Score of 94.2, placing it on the Pareto frontier. Ablation studies revealed that Fourier encoding is the single most critical component: removing it drops performance by 30.7 percentage points and fails the MEP gate on 14% of tasks. Surprisingly, global EWC regulation degrades performance (NES = -4.2), while topology-local EWC reduces the penalty but doesn't eliminate it. Removing both EWC and Hebbian updates yields MPCS_EFFICIENT, which improves performance by 0.6 pp at 4.7x lower compute cost (127 vs. 602 minutes), validating the Pareto frontier as a practical model-compression guide.