Scholarlink

<p>Spiking neural networks (SNNs) have garnered attention for their energy efficiency in neuromorphic computing. However, existing spiking ResNets often suffer from limited representational capacity and scalability issues, primarily due to information loss from sparse binary spikes and the incompatibility of certain effective modules designed for artificial neural networks (ANNs). Furthermore, the multi-time-step simulation of SNNs entails greater computational resource demands compared to ANNs, relying on GPU-based supercomputing platforms, which also hinders the scaling of SNNs to larger architectures. This study designs an enhanced Spiking ResNet architecture, termed E-SResNet, with three lightweight, plug-and-play modules to enhance the representational capacity, while simultaneously maintaining manageable model complexity to achieve scalability. Firstly, a SpaceToDepth stem module directly restructures spatial pixels into the channel dimension to efficiently reduce the input resolution, meeting the demand for higher quality in early feature extraction due to sparse spiking activity. Secondly, a Spiking Anti-Alias (SAA) downsampling module employs a spike-driven low-pass filter and adjusts padding to mitigate spatio-temporal aliasing artifacts during downsampling, while optimizing gradient backpropagation for model efficiency and stability. Thirdly, a Spiking Squeeze-and-Excitation (SSE) module dynamically recalibrates channel-wise spike responses through spike activation, capturing channel-temporal features, while controlling model complexity with a reduction ratio. Extensive experiments on static datasets (CIFAR10/100, ImageNet-1K) and neuromorphic benchmarks (CIFAR10-DVS, DVS128 Gesture) demonstrate that E-SResNet achieves state-of-the-art performance with negligible additional parameter overhead, meeting high-performance computing demands, underscoring the effectiveness of spike-driven architectural enhancements in improving the representation capacity and scalability of deep spiking ResNets. The code is available at <a href="https://github.com/BioCenter-SHU/E-SResNet">https://github.com/BioCenter-SHU/E-SResNet</a></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Enhancing the representational capacity and scalability of ResNet architectures based on spiking neural networks

  • 期刊论文

  • 发表日期：

    2025年11月19日

查看全文
  • Qi Jin,
  • Jiang Xie

摘要

Spiking neural networks (SNNs) have garnered attention for their energy efficiency in neuromorphic computing. However, existing spiking ResNets often suffer from limited representational capacity and scalability issues, primarily due to information loss from sparse binary spikes and the incompatibility of certain effective modules designed for artificial neural networks (ANNs). Furthermore, the multi-time-step simulation of SNNs entails greater computational resource demands compared to ANNs, relying on GPU-based supercomputing platforms, which also hinders the scaling of SNNs to larger architectures. This study designs an enhanced Spiking ResNet architecture, termed E-SResNet, with three lightweight, plug-and-play modules to enhance the representational capacity, while simultaneously maintaining manageable model complexity to achieve scalability. Firstly, a SpaceToDepth stem module directly restructures spatial pixels into the channel dimension to efficiently reduce the input resolution, meeting the demand for higher quality in early feature extraction due to sparse spiking activity. Secondly, a Spiking Anti-Alias (SAA) downsampling module employs a spike-driven low-pass filter and adjusts padding to mitigate spatio-temporal aliasing artifacts during downsampling, while optimizing gradient backpropagation for model efficiency and stability. Thirdly, a Spiking Squeeze-and-Excitation (SSE) module dynamically recalibrates channel-wise spike responses through spike activation, capturing channel-temporal features, while controlling model complexity with a reduction ratio. Extensive experiments on static datasets (CIFAR10/100, ImageNet-1K) and neuromorphic benchmarks (CIFAR10-DVS, DVS128 Gesture) demonstrate that E-SResNet achieves state-of-the-art performance with negligible additional parameter overhead, meeting high-performance computing demands, underscoring the effectiveness of spike-driven architectural enhancements in improving the representation capacity and scalability of deep spiking ResNets. The code is available at https://github.com/BioCenter-SHU/E-SResNet