Scholarlink

<p>Spiking Neural Networks (SNNs) offer a low-power alternative to traditional Artificial Neural Networks (ANNs). However, their performance often lags behind due to challenges in training and the discrete nature of spike-based signal transmission. In this paper, we propose the Spiking ST-Former, a novel spiking transformer architecture that addresses these limitations by incorporating a Spatio-Temporal Spiking Self-Attention (STSA) mechanism. STSA enables efficient joint modeling of spatial and temporal information through a Temporal Processing Module (TPM), without increasing the number of model parameters. Experimental results on both static and neuromorphic datasets, including CIFAR-10/100, ImageNet, CIFAR10-DVS, and DVS128, demonstrate that STSA consistently outperforms spatial-only attention mechanisms and achieves competitive performance. These findings highlight the potential of STSA in enhancing SNN performance and advancing the development of spiking transformers. Code will be made publicly available at <a href="https://github.com/Bennie123-byte/Spiking-ST-former">https://github.com/Bennie123-byte/Spiking-ST-former</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

Spiking ST-former: enhancing spatio-temporal modeling in spiking transformers via integrated self-attention mechanisms

  • 期刊论文

  • 发表日期：

    2025年9月9日

查看全文
  • Yijun Liu,
  • Bin Liu,
  • Wujian Ye,
  • Guoliang Tan,
  • Youfeng Cui

摘要

Spiking Neural Networks (SNNs) offer a low-power alternative to traditional Artificial Neural Networks (ANNs). However, their performance often lags behind due to challenges in training and the discrete nature of spike-based signal transmission. In this paper, we propose the Spiking ST-Former, a novel spiking transformer architecture that addresses these limitations by incorporating a Spatio-Temporal Spiking Self-Attention (STSA) mechanism. STSA enables efficient joint modeling of spatial and temporal information through a Temporal Processing Module (TPM), without increasing the number of model parameters. Experimental results on both static and neuromorphic datasets, including CIFAR-10/100, ImageNet, CIFAR10-DVS, and DVS128, demonstrate that STSA consistently outperforms spatial-only attention mechanisms and achieves competitive performance. These findings highlight the potential of STSA in enhancing SNN performance and advancing the development of spiking transformers. Code will be made publicly available at https://github.com/Bennie123-byte/Spiking-ST-former.