Research on Anti-Interference Performance of Spiking Neural Network Under Network Connection Damage

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Bionics

Background:

• Memristors offer combined storage and computing, ideal for optimizing artificial intelligence (AI) and neural networks.
• Spiking neural networks (SNNs) exhibit inherent resistance to various interferences through synaptic plasticity.
• The anti-interference capabilities of SNNs are crucial for advancing electromagnetic protection bionics.

Purpose of the Study:

• To construct and evaluate spiking neural network (SNN) models against traditional artificial neural networks (ANNs).
• To investigate the anti-interference performance of SNNs and ANNs under simulated network connection damage.
• To compare the robustness of VGG-SNN and FCNN-SNN models using real-world and benchmark datasets.

Main Methods:

• Developed two SNN models (VGG-SNN, FCNN-SNN) using the Leaky Integrate-and-Fire (LIF) neuron model.
• Employed a pruning algorithm to simulate network connection damage during training.
• Evaluated performance on millimeter wave radar human motion and MNIST datasets, comparing SNNs with ANNs at 30% sparsity.

Main Results:

• On the human motion dataset, SNN accuracy dropped by 5.83% vs. 18.71% for ANN at 30% sparsity.
• On the MNIST dataset, SNN accuracy decreased by 3.91% vs. 10.13% for ANN at 30% sparsity.
• SNNs demonstrated significantly better resilience to network damage than ANNs.

Conclusions:

• SNNs possess distinct anti-interference advantages over ANNs under equivalent network damage.
• SNNs offer more stable and superior performance in information processing and pattern recognition.
• Network structure, encoding, and learning algorithms critically influence SNN anti-interference performance.