CogniSNN: Enabling neuron-expandability, pathway-reusability, and dynamic-configurability in spiking neural networks.

Abstract

Spiking neural networks (SNNs) hold promise for bridging artificial intelligence and computational neuroscience. However, mainstream SNNs predominantly adopt the rigid, chain-like architectures of artificial neural networks (ANNs), ignoring key structural characteristics. Biological neurons are stochastically interconnected, forming complex neural pathways that exhibit Neuron-Expandability, Pathway-Reusability, and Dynamic-Configurability. To explicitly model these characteristics, we introduce Cognition-aware SNN (CogniSNN), a novel paradigm grounded in Random Graph Architecture (RGA). Furthermore, we address deep-network degradation and dimensional mismatch via a purely spiking residual mechanism and an adaptive pooling strategy. Then, we design a Key Pathway-based Learning without Forgetting (KP-LwF) approach that selectively reuses structural pathways for efficient multi-task continual learning. Finally, we also propose a Dynamic Growth Learning (DGL) algorithm, enabling neurons and synapses to evolve dynamically along the temporal dimension. Extensive experiments on neuromorphic datasets and Tiny-ImageNet demonstrate that CogniSNN achieves state-of-the-art performance. Notably, it reaches 80.64% accuracy on N-Caltech101 with only 5 timesteps, outperforming SSNN by 2.67%. Beyond static accuracy, CogniSNN exhibits superior continual learning capabilities, enhanced noise robustness, and mitigated fixed-timestep constraints. This work demonstrates the potential of SNNs with random graph structures in advancing brain-inspired intelligence and lays the foundation for their practical application on neuromorphic hardware. The code is available at https://github.com/Yongsheng124/CogniSNN.