Community-aware sparse topology design for efficient spiking neural networks.

Area of Science:

• Neuromorphic Computing
• Artificial Intelligence
• Network Science

Background:

• Spiking Neural Networks (SNNs) are crucial for energy-efficient neuromorphic computing.
• Current SNNs often adapt dense Artificial Neural Network (ANN) structures, neglecting topology's role in learning.
• Optimizing SNN topology is vital for enhancing performance and efficiency.

Purpose of the Study:

• To introduce a community-aware sparse topology design framework for graph-based SNNs.
• To investigate the impact of various community detection algorithms on SNN learning dynamics.
• To provide guidelines for designing efficient and accurate SNNs through topology optimization.

Main Methods:

• Employed seven community detection algorithms (KMeans, Spectral Clustering, Fast Greedy, Louvain, Leiden, Infomap, Small-World) to structure SNN topologies.
• Systematically compared convergence speed, classification accuracy, and energy consumption under controlled conditions (64 neurons, 92% sparsity).
• Evaluated performance on MNIST and CIFAR-10 datasets.

Main Results:

• Community-driven topologies exhibited dramatically faster convergence (27-44 epochs vs. 100-300 epochs) compared to conventional SNNs.
• Dataset-dependent accuracy trade-offs observed: Infomap excelled on MNIST (99.67%), while Louvain, KMeans, and Small-World performed better on CIFAR-10 (≈92.96%).
• Sparse modular architectures maintained low inference energy (≈1.2 mJ/sample) with higher sparsity and structured connectivity, halving energy per neuron.

Conclusions:

• Network topology critically influences SNN learning efficiency, accuracy, and energy consumption, challenging assumptions based solely on size or sparsity.
• The proposed framework offers practical, dataset-aware guidelines for designing neuromorphic systems.
• Optimizing community structure in SNNs is essential for advancing energy-efficient AI.