Problem

Traditional Spiking Neural Networks (SNNs) primarily adopt rigid, chain-like hierarchical architectures inspired by Artificial Neural Networks (ANNs), often overlooking crucial structural characteristics of biological brains, such as stochastic interconnections and complex neural pathways. This architectural rigidity leads to significant challenges, including limitations in Neuron-Expandability (difficulty in handling deep networks due to degradation and spatial dimension mismatches), Pathway-Reusability (poor performance in multi-task scenarios and catastrophic forgetting), and Dynamic-Configurability (lack of robustness against interference and fixed-timestep constraints for hardware deployment). Existing SNN residual mechanisms often introduce non-spiking, floating-point operations, contradicting the event-driven nature of SNNs, while current continual learning methods lack the granularity to adaptively select specific pathways for new knowledge acquisition. Furthermore, attempts at dynamic network adaptation typically rely on inefficient pruning methods.

Method

The authors propose Cognition-aware SNN (CogniSNN), a new SNN paradigm built upon Random Graph Architectures (RGA), which model the stochastic interconnections found in biological brains. The RGA is constructed as a Directed Acyclic Graph (DAG) using WS and ER network generators. To address the identified problems, CogniSNN integrates three key mechanisms:

• Neuron-Expandability: An OR Gate residual mechanism replaces traditional arithmetic addition with a logical OR operation, ensuring strictly binary spike propagation, preventing unbounded value accumulation, and allowing for mathematically sound identity mapping. This is complemented by an Adaptive Pooling (AP) strategy that dynamically resizes feature maps from different predecessor nodes to resolve spatial dimension mismatches in the random graph.
• Pathway-Reusability: The Key Pathway-based Learning without Forgetting (KP-LwF) algorithm enables efficient multi-task transfer by selectively reusing neural pathways. This algorithm leverages Pathway Betweenness Centrality (BC) to identify critical pathways. For similar tasks, high-BC pathways (encoding generalizable features) are fine-tuned, while for dissimilar tasks, low-BC pathways (encoding marginal features) are selected to minimize interference with established knowledge.
• Dynamic-Configurability: A Dynamic Growth Learning (DGL) algorithm simulates the progressive expansion of biological neural networks. During training, the network operates on a progressively expanding set of neural pathways over timesteps. During inference, all pathways are utilized, including those partially trained, which significantly enhances robustness against noise and frame loss, and improves deployment flexibility by mitigating fixed-timestep constraints on neuromorphic hardware.

Results

Extensive experiments demonstrate CogniSNN's effectiveness across various tasks and scenarios:

• Classification Performance: CogniSNN achieves performance comparable to, and often surpasses, current state-of-the-art SNNs on neuromorphic datasets (DVS-Gesture, CIFAR10-DVS, N-Caltech101) and static datasets (Tiny-ImageNet). It shows robust performance even in low-latency settings, indicating superior spatial feature extraction capabilities.
• Continual Learning: The KP-LwF algorithm significantly outperforms a standard Learning without Forgetting (LwF) baseline by nearly 10%. Specifically, low-BC pathways are more effective for dissimilar tasks (e.g., CIFAR100 to MNIST, 3.1% advantage), while high-BC pathways are better for similar tasks (e.g., CIFAR100 to CIFAR10, 1.8% advantage).
• Anti-Interference Robustness: The DGL algorithm provides superior robustness against various types of random noise (Salt-and-Pepper noise: up to 16.38% advantage; Poisson noise: ~9.79% advantage) and frame-loss (10.10% advantage) compared to a static CogniSNN baseline.
• Deployment Flexibility: DGL-trained CogniSNN exhibits significantly better performance retention when inference timesteps are reduced, achieving up to 17.4% performance retention over the baseline on DVS-Gesture with a single timestep, and 16.3% on CIFAR10-DVS with two timesteps.
• Ablation Studies: Random graph structures consistently yield over 2% higher accuracy than chain-like architectures, attributed to implicit skip connections. The OR Gate mechanism provides a 9.0% accuracy improvement over no skip connections and a 0.4% improvement over ADD-based residuals, while maintaining energy efficiency and binary operations.

Implications

This work highlights the immense potential of incorporating random graph structures into Spiking Neural Networks for advancing brain-inspired artificial intelligence. By explicitly modeling Neuron-Expandability, Pathway-Reusability, and Dynamic-Configurability, CogniSNN offers a robust and flexible framework that overcomes critical limitations of conventional SNNs. The proposed OR Gate, Adaptive Pooling, KP-LwF, and DGL algorithms provide practical solutions for network degradation, catastrophic forgetting, and fixed-timestep constraints, paving the way for more robust, adaptable, and continually learning SNNs suitable for real-world multi-task applications and efficient deployment on neuromorphic hardware.