A Spatial-Channel-Temporal-Fused Attention for Spiking Neural Networks.

Area of Science:

• Neuroscience and Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Spiking neural networks (SNNs) mimic brain computation for spatiotemporal data processing.
• Visual attention is crucial for perception but underutilized in SNNs.
• Predictive attentional remapping in biology inspires new SNN mechanisms.

Purpose of the Study:

• To propose and evaluate a novel spatial-channel-temporal-fused attention (SCTFA) module for SNNs.
• To enhance SNNs' ability to capture salient information using historical context.
• To improve SNN performance in event-based vision tasks.

Main Methods:

• Development of the spatial-channel-temporal-fused attention (SCTFA) module.
• Integration of SCTFA into a Spiking Neural Network (SNN) architecture.
• Systematic evaluation on DVS Gesture, SL-Animals-DVS, and MNIST-DVS datasets.

Main Results:

• The SCTFA-SNN significantly outperformed baseline SNNs and SNNs with degraded attention modules.
• The proposed model achieved competitive accuracy with existing state-of-the-art methods.
• SCTFA-SNN demonstrated robustness to noise and stability with incomplete data.

Conclusions:

• Incorporating cognitive mechanisms like visual attention can significantly enhance SNN capabilities.
• The SCTFA module offers a promising approach for improving SNN performance in complex tasks.
• The SCTFA-SNN presents an efficient and robust solution for event-based visual processing.