SD²-SNN: Self-distillation and structural decomposition framework for SNNs in continual learning.

Abstract

Catastrophic forgetting remains a fundamental obstacle for Artificial Neural Networks (ANNs) in continual learning. Although existing ANN-based methods can alleviate forgetting, they often introduce additional overhead and are less suitable for low-power deployment. Spiking Neural Networks (SNNs) provide an energy-efficient alternative, yet current SNN continual-learning methods still rely on external teacher models or task identifiers and lack a unified mechanism to jointly stabilize spike-level representations and synaptic parameters over long task sequences. To mitigate both limitations, this study proposes SD²-SNN, a framework employing Self-Distillation and Structural Decomposition to enhance knowledge retention without external supervision. To mitigate forgetting, SD²-SNN integrates two synergistic mechanisms: it utilizes internal self-distillation to anchor decision boundaries by aligning spike-rate distributions with prior states, and concurrently implements structural weight decomposition to decouple parameters into a stable shared base and a dynamic task-specific component. This approach effectively balances plasticity and stability while leveraging inherent SNNs sparsity. Experiments on multiple continual learning benchmarks demonstrate that SD²-SNN achieves strong and stable performance across both image-based and event-based settings. In particular, it attains 57.98% and 47.78% CIL accuracy on Split-CIFAR100 under 10-step and 20-step protocols, 71.5%/40.4% TIL/CIL on Tiny-ImageNet, and 90.3%/62.9% TIL/CIL on DVS128 Gesture.