Scholarlink

<p>Spiking Neural Networks (SNN) provide effective bio-inspired solutions for computer vision applications within the domain of neuromorphic computing. However, present SNNs face two limitations 1) their asynchronous and non-differentiable spike activity makes training difficult, and 2) they struggle to extract temporal data from images over longer timesteps, resulting in performance degradation. To address these limitations, this paper proposes two different SNN implementation approaches based on Leaky Integrate-and-Fire (LIF) neurons. In addition, an ablation study is conducted by experimenting with various hyperparameters and optimizers to identify an optimal SNN configuration. Furthermore, a comparative analysis of both approaches is performed using the Two-hands Indian Sign Language (ISL) alphabet dataset. The experimental results demonstrate that: 1) the proposed SNN with the NengoDL framework and RMSProp optimizer achieves the highest recognition accuracy of 99.8% on the ISL dataset, thus surpassing existing state-of-the-art methods. 2) The model has a memory footprint of 9.05 MB, making it feasible for deployment on neuromorphic hardware for real-time applications.</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

A comprehensive performance analysis of SNN-based neuromorphic architectures for hand gesture recognition

  • 期刊论文

  • 发表日期：

    2026年1月14日

查看全文
  • Kartik Khandelwal,
  • Viren Sharma,
  • Rashi Goyal,
  • Mohita Jaiswal,
  • Jawar Singh,
  • Abhishek Sharma

摘要

Spiking Neural Networks (SNN) provide effective bio-inspired solutions for computer vision applications within the domain of neuromorphic computing. However, present SNNs face two limitations 1) their asynchronous and non-differentiable spike activity makes training difficult, and 2) they struggle to extract temporal data from images over longer timesteps, resulting in performance degradation. To address these limitations, this paper proposes two different SNN implementation approaches based on Leaky Integrate-and-Fire (LIF) neurons. In addition, an ablation study is conducted by experimenting with various hyperparameters and optimizers to identify an optimal SNN configuration. Furthermore, a comparative analysis of both approaches is performed using the Two-hands Indian Sign Language (ISL) alphabet dataset. The experimental results demonstrate that: 1) the proposed SNN with the NengoDL framework and RMSProp optimizer achieves the highest recognition accuracy of 99.8% on the ISL dataset, thus surpassing existing state-of-the-art methods. 2) The model has a memory footprint of 9.05 MB, making it feasible for deployment on neuromorphic hardware for real-time applications.