cognisnn-random-graph-architecture

name	cognisnn-random-graph-architecture
description	CogniSNN: Enabling Neuron-Expandability, Pathway-Reusability, and Dynamic-Configurability with Random Graph Architectures in Spiking Neural Networks
category	neuroscience
tags	["spiking neural network","random graph architecture","neuron-expandability","pathway-reusability","dynamic-configurability"]
arxiv_id	2512.11743

CogniSNN: Random Graph Architecture for Spiking Neural Networks

Context

This skill implements the CogniSNN framework introduced in arXiv:2512.11743, which addresses the limitations of traditional spiking neural networks (SNNs) that rigidly follow chain-like hierarchical architectures of traditional ANNs. The paper introduces Random Graph Architecture (RGA) to incorporate three key biological properties: Neuron-Expandability, Pathway-Reusability, and Dynamic-Configurability.

Core Methodology

1. Random Graph Architecture (RGA): Replace rigid hierarchical connections with stochastic interconnections mimicking biological neural networks
2. Improved Pure Spiking Residual Mechanism: Address network degradation and dimensional mismatch in deep pathways
3. Adaptive Pooling Strategy: Complement the residual mechanism for stable feature extraction
4. Key Pathway-based Learning without Forgetting (KP-LwF): Selectively reuse critical neural pathways while retaining historical knowledge for efficient multi-task transfer
5. Dynamic Growth Learning (DGL) Algorithm: Allow neurons and synapses to grow dynamically along the internal temporal dimension

Implementation Steps

1. Network Initialization with RGA:

   • Generate random connectivity patterns using Erdős–Rényi or Watts-Strogatz models
   • Ensure biological plausibility through connection probability tuning
   • Implement both excitatory and inhibitory connections with realistic ratios

2. Spiking Neuron Model Selection:

   • Choose appropriate spiking neuron model (LIF, Izhikevich, etc.)
   • Implement membrane potential dynamics and spike generation
   • Configure refractory periods and threshold dynamics

3. Residual Connection Implementation:

   • Design identity mapping connections that bypass problematic layers
   • Implement spike-based residual addition operations
   • Add adaptive pooling layers to complement residual connections

4. KP-LwF Learning Mechanism:

   • Identify critical pathways for task performance
   • Implement selective pathway reuse during task switching
   • Develop forgetting mitigation strategies for preserved knowledge

5. Dynamic Growth Algorithm:

   • Monitor network activity and performance metrics
   • Trigger neuron/synapse addition when performance plateaus
   • Integrate new components through structured growth protocols

6. Training Procedure:

   • Implement event-driven simulation for efficiency
   • Apply spike-timing dependent plasticity (STDP) or surrogate gradients
   • Validate on neuromorphic datasets (DVS-Gesture, CIFAR10-DVS) and Tiny-ImageNet

Configuration Parameters

• Connection probability (p): 0.1-0.3 for sparse connectivity
• Neuron growth rate: Adaptive based on performance metrics
• Pathway reuse threshold: Task-specific similarity measure
• Temporal dimension growth rate: Configurable based on task complexity

Verification

• Performance comparison with state-of-the-art SNNs on benchmark datasets
• Analysis of pathway reusability across sequential tasks
• Evaluation of dynamic growth impact on network robustness
• Verification of biological plausibility metrics

Activation Keywords

cognisnn, random graph architecture, spiking neural network, neuron-expandability, pathway-reusability, dynamic-configurability, KP-LwF, DGL

References

• arXiv:2512.11743: CogniSNN: Enabling Neuron-Expandability, Pathway-Reusability, and Dynamic-Configurability with Random Graph Architectures in Spiking Neural Networks