SGLFormer: Spiking Global-Local-Fusion Transformer with high performance

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Computer Vision

Background:

• Spiking Neural Networks (SNNs) are biologically inspired models known for low energy consumption and event-driven processing.
• Current SNNs face performance limitations despite their potential.
• The brain's complex architecture inspires advanced information processing capabilities.

Purpose of the Study:

• To introduce a novel Spiking Neural Network architecture, the Spiking Global-Local-Fusion Transformer (SGLFormer).
• To significantly improve the performance of SNNs in computer vision tasks.
• To address limitations in existing SNN architectures, particularly concerning gradient backpropagation.

Main Methods:

• Developed the Spiking Global-Local-Fusion Transformer (SGLFormer) by integrating transformer and convolution structures.
• Introduced a novel Maxpooling module to address inaccurate gradient backpropagation issues in SNNs.
• Utilized spatio-temporal blocks (STB) in the classification head for enhanced feature aggregation.

Main Results:

• SGLFormer demonstrated superior performance on static datasets (CIFAR10/100, ImageNet) and dynamic vision sensor (DVS) datasets (CIFAR10-DVS, DVS128-Gesture).
• Achieved a top-1 accuracy of 83.73% on ImageNet with 64 M parameters, surpassing current state-of-the-art SNNs by 6.66%.
• Outperformed existing SNNs in directly trained benchmarks.

Conclusions:

• The SGLFormer architecture offers a significant advancement in SNN performance for computer vision.
• The novel design effectively processes information at both global and local scales.
• SGLFormer shows promise for supporting a wider range of future computer vision applications.