Adaptive Sparse Structural Evolution for Continual Learning in Spiking Neural Networks

Lifelong learning enables intelligent systems to acquire knowledge incrementally while preserving previously learned information—a hallmark of human cognition. However, most Spiking Neural Network (SNN) models lack the critical biological mechanisms that support this capacity, particularly sparse circuit reconfiguration and context-aware gating observed in the brain. Inspired by neuroscientific insights into synaptic rewiring and task-differentiated cortical modulation, we propose a novel framework for sparse structural continual learning in SNNs. Our approach integrates two core innovations: (1) a context-aware gating mechanism based on logistic regression that detects task-specific representations from spatio-temporal spike activity, and (2) an iterative sparse structure learning module that dynamically prunes and regrows task-specific connections in a biologically plausible manner. Together, these components facilitate efficient representation segregation, reduce catastrophic forgetting, and improve parameter efficiency. Experimental results on class-incremental CIFAR-100 benchmarks show significant gains in both continual learning performance and computational efficiency over state-of-the-art SNN baselines, which use only half the number of parameters, with up to 1.11% improvement in average accuracy and up to 2.67% increase in last accuracy. This work offers a biologically grounded pathway to scalable and adaptive lifelong learning in neuromorphic AI systems.