| name | cognisnn-random-graph-snn |
| description | CogniSNN: Cognition-aware Spiking Neural Network with Random Graph Architecture enabling neuron-expandability, pathway-reusability, and dynamic-configurability. Uses Key Pathway-based Learning (KP-LwF) for multi-task transfer and Dynamic Growth Learning (DGL) algorithm for temporal dimension growth. Achieves SOTA on neuromorphic datasets. Keywords: SNN architecture, random graph, pathway reusability, dynamic growth, neuromorphic hardware, continual learning, brain-inspired AI. |
| tags | ["spiking-neural-network","neuromorphic","random-graph-architecture","continual-learning","dynamic-growth","pathway-reusability","cognition-aware"] |
CogniSNN: Cognition-aware SNN with Random Graph Architecture
Paper Information
- arXiv ID: 2512.11743
- Title: CogniSNN: Enabling Neuron-Expandability, Pathway-Reusability, and Dynamic-Configurability with Random Graph Architectures in Spiking Neural Networks
- Authors: Yongsheng Huang, Peibo Duan, Yujie Wu, Kai Sun, Zhipeng Liu, Changsheng Zhang, Bin Zhang, Mingkun Xu
- Submission Date: 2025-12-12
- Categories: cs.NE (Neural and Evolutionary Computing), cs.AI (Artificial Intelligence)
- DOI: https://doi.org/10.48550/arXiv.2512.11743
Core Innovation
CogniSNN introduces Random Graph Architecture (RGA) to SNNs, breaking from traditional rigid chain-like hierarchical structures to mimic biological neural connectivity patterns.
Three Key Properties
- Neuron-Expandability: Network can dynamically add neurons without disrupting existing structure
- Pathway-Reusability: Critical neural pathways can be selectively reused across tasks
- Dynamic-Configurability: Network topology can adapt during training and deployment
Key Technical Contributions
1. Improved Pure Spiking Residual Mechanism
- Addresses network degradation in deep pathways
- Implements adaptive pooling strategy to handle dimensional mismatch
- Maintains spike-based computation throughout (no conversion to ANN)
2. Key Pathway-based Learning without Forgetting (KP-LwF)
- Selectively reuses critical neural pathways for multi-task transfer
- Retains historical knowledge during new task learning
- Enables efficient continual learning without catastrophic forgetting
- Identifies "key pathways" based on importance metrics
3. Dynamic Growth Learning (DGL) Algorithm
- Allows neurons and synapses to grow dynamically along temporal dimension
- Adaptively expands network capacity based on task complexity
- Mitigates fixed-timestep constraints on neuromorphic hardware
- Improves robustness against interference
Random Graph Architecture Design
Structural Properties
- Stochastic Connectivity: Mimics biological brain's random interconnections
- Multi-path Routing: Information flows through multiple parallel pathways
- Flexible Depth: Pathway length adapts based on computational needs
- Sparse Connections: Reduces synaptic overhead while maintaining expressivity
Implementation Strategy
class RandomGraphSNN:
def __init__(self, neuron_pool_size, connectivity_probability):
self.neurons = NeuronPool(neuron_pool_size)
self.random_connect(connectivity_probability)
def random_connect(self, prob):
for src in self.neurons:
for dst in self.neurons:
if random.random() < prob:
create_synapse(src, dst)
Performance Results
Neuromorphic Datasets
- Comparable or surpassing state-of-the-art SNN performance
- Demonstrates effectiveness on DVS Gesture, CIFAR10-DVS, N-Caltech101
Tiny-ImageNet
- Successfully scales to larger-scale vision tasks
- Maintains spike-based computation efficiency
Key Metrics
- Energy Efficiency: Leverages SNN's sparse activation patterns
- Multi-task Transfer: KP-LwF enables seamless knowledge transfer
- Hardware Compatibility: Addresses neuromorphic deployment constraints
Biological Inspiration
CogniSNN draws from three key biological principles:
-
Stochastic Neural Connectivity
- Brain neurons connect probabilistically, not in rigid layers
- Random graphs capture this biological variability
-
Neural Pathway Reuse
- Brain reuses established pathways for related tasks
- KP-LwF mimics this efficient knowledge reuse mechanism
-
Dynamic Neural Growth
- Biological networks can grow during learning (neurogenesis)
- DGL algorithm simulates this adaptive expansion
Practical Applications
Neuromorphic Hardware Deployment
- Loihi/Loihi 2: Intel neuromorphic chips benefit from dynamic growth
- SpiNNaker: ARM-based neuromorphic platform compatible with RGA
- TrueNorth: IBM's neurosynaptic processor supports pathway reuse
Continual Learning Scenarios
- Robotics: Sequential task learning without retraining
- IoT Edge Devices: Memory-efficient multi-task models
- Autonomous Systems: Adaptive learning during operation
Implementation Guidelines
Step 1: Random Graph Construction
connectivity_prob = 0.3
graph = RandomGraph(neuron_count=1000, edge_prob=connectivity_prob)
Step 2: Pure Spiking Residual Block
class SpikingResidualBlock:
def __init__(self, channels):
self.conv1 = SpikingConv(channels)
self.conv2 = SpikingConv(channels)
self.pool = AdaptiveSpikePool()
def forward(self, x):
identity = x
out = self.conv1(x)
out = self.conv2(out)
out = self.pool(out + identity)
return out
Step 3: Key Pathway Identification
def identify_key_pathways(model, task_history):
importance = compute_pathway_importance(model, task_history)
key_pathways = threshold_selection(importance, top_k=0.3)
return key_pathways
Step 4: Dynamic Growth
class DynamicGrowthController:
def __init__(self, growth_threshold=0.8):
self.threshold = growth_threshold
def should_grow(self, current_performance):
if current_performance < self.threshold:
return True
return False
def grow_neurons(self, model, num_new_neurons):
model.add_neurons(num_new_neurons, target_region='high_activity')
Key Insights
Architecture Philosophy
- Reject Layer-by-Layer Dogma: Biological brain doesn't use rigid sequential layers
- Embrace Randomness: Stochastic connectivity enables richer representations
- Enable Flexibility: Network should adapt structure during learning
Learning Mechanisms
- Pathway-Level Knowledge: Store knowledge in pathways, not just weights
- Selective Reuse: Identify and preserve high-value pathways across tasks
- Dynamic Expansion: Grow capacity when performance plateaus
Hardware Considerations
- Fixed-Timestep Problem: Traditional SNNs constrained by predetermined simulation time
- Dynamic Timestep: DGL allows variable simulation duration
- Memory Efficiency: Random graphs reduce redundant weight storage
Comparison with Traditional SNNs
| Feature | Traditional SNN | CogniSNN |
|---|
| Architecture | Sequential layers | Random graph |
| Connectivity | Dense/rigid | Sparse/stochastic |
| Multi-task | Separate models | Shared pathways |
| Capacity | Fixed neurons | Dynamic growth |
| Learning | Single task | Continual learning |
Limitations and Considerations
- Graph Search Cost: Finding key pathways requires additional computation
- Hyperparameter Tuning: Connectivity probability needs careful selection
- Hardware Mapping: Random graphs may not align with structured neuromorphic chips
- Training Complexity: Multiple mechanisms (KP-LwF + DGL) increase optimization difficulty
Future Directions
Research Opportunities
- Graph Topology Optimization: Learn connectivity patterns instead of random initialization
- Hardware-Aware Graph Design: Co-design random graphs for specific neuromorphic platforms
- Biological Validation: Compare CogniSNN pathways with real cortical connectivity maps
Technical Extensions
- Hierarchical Random Graphs: Combine local and global stochastic connectivity
- Temporal Randomness: Dynamic connectivity changes during simulation
- Multi-modal Integration: RGA for vision-language SNN architectures
Code and Resources
Implementation Status
- Paper provides detailed algorithm descriptions
- No official code release at time of analysis
- Potential for implementation using existing SNN frameworks (SpikingJelly, Norse)
Related Work
- Traditional SNN architectures: Surrogate gradient methods
- Graph Neural Networks: Structured graph learning
- Continual Learning: Elastic Weight Consolidation, Progressive Networks
Citation
@article{huang2025cognisnn,
title={CogniSNN: Enabling Neuron-Expandability, Pathway-Reusability, and Dynamic-Configurability with Random Graph Architectures in Spiking Neural Networks},
author={Huang, Yongsheng and Duan, Peibo and Wu, Yujie and Sun, Kai and Liu, Zhipeng and Zhang, Changsheng and Zhang, Bin and Xu, Mingkun},
journal={arXiv preprint arXiv:2512.11743},
year={2025}
}
Activation Triggers
Use this skill when working on:
- SNN architecture design: Breaking from traditional layer structures
- Continual learning: Multi-task knowledge retention in SNNs
- Neuromorphic deployment: Hardware-aware SNN optimization
- Biological inspiration: Brain-inspired connectivity patterns
- Dynamic network growth: Adaptive capacity during learning
- Pathway reuse: Efficient knowledge transfer mechanisms
Keywords: cognisnn, random graph snn, pathway reusability, dynamic growth learning, kp-lwf, neuron expandability, spiking residual, neuromorphic hardware, continual learning snn, brain-inspired architecture, stochastic connectivity, adaptive pooling, neural pathway, temporal growth