Scholarlink

<p>In present study, an energy efficient higher-order leaky integrate-and-fire (HLIF) neurone model is proposed to enhance the temporal learning capability and classification performance compared to first-order LIF leaky integrate-and-fire, generalized LIF (GLIF) and relaxation LIF neuron models. The theoretical formulations of the proposed neurone are articulated through an analysis of membrane potential dynamics and the rate of membrane decay (RoD), offering insights into temporal retention, membrane persistence, and spike propagation characteristics of HLIF neurone model. Furthermore, spike encoding framework is employed via spike, image, and raster formation, along with temporal membrane state visualisation on benchmark datasets (i.e. MNIST, Fashion-MNIST, CIFAR-10, and CIFAR-100), to assess the temporal dynamics of the encoded spike streams. Subsequently, an HLIF-based SNN architecture is also developed and assessed for image classification on benchmark datasets. Experimental results indicate that the proposed HLIF neurone attains a remarkable 96% enhancement in membrane decay regulation across 30 simulation time-steps compared to first-order LIF neuron models. This advancement leads to increased spike sparsity, greater energy efficiency, and optimised utilisation of computational resources. The proposed model also demonstrates significant enhancements in energy efficiency, achieving improvements of up to 25%, 20%, and 26% on the MNIST, CIFAR-10, and CIFAR-100 datasets, respectively. Additionally, resource efficiency improvements are notable, reaching up to 27%, 34%, 24%, and 38% on the MNIST, Fashion-MNIST, CIFAR-10, and CIFAR-100 datasets, respectively. Moreover, the proposed HLIF neurone significantly improves temporal processing, adaptive inference capabilities, energy efficiency, and classification reliability, positioning it as a promising method for optimising deep spiking neural networks.</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

An energy efficient higher order leaky neuron model with spike encoding

  • 期刊论文

  • 发表日期：

    2026年8月3日

查看全文
  • J. Mukhopadhyay,
  • B. Manna,
  • S. Mandal,
  • M. Rakshit,
  • J. K. Rakshit,
  • D. Acharyya

摘要

In present study, an energy efficient higher-order leaky integrate-and-fire (HLIF) neurone model is proposed to enhance the temporal learning capability and classification performance compared to first-order LIF leaky integrate-and-fire, generalized LIF (GLIF) and relaxation LIF neuron models. The theoretical formulations of the proposed neurone are articulated through an analysis of membrane potential dynamics and the rate of membrane decay (RoD), offering insights into temporal retention, membrane persistence, and spike propagation characteristics of HLIF neurone model. Furthermore, spike encoding framework is employed via spike, image, and raster formation, along with temporal membrane state visualisation on benchmark datasets (i.e. MNIST, Fashion-MNIST, CIFAR-10, and CIFAR-100), to assess the temporal dynamics of the encoded spike streams. Subsequently, an HLIF-based SNN architecture is also developed and assessed for image classification on benchmark datasets. Experimental results indicate that the proposed HLIF neurone attains a remarkable 96% enhancement in membrane decay regulation across 30 simulation time-steps compared to first-order LIF neuron models. This advancement leads to increased spike sparsity, greater energy efficiency, and optimised utilisation of computational resources. The proposed model also demonstrates significant enhancements in energy efficiency, achieving improvements of up to 25%, 20%, and 26% on the MNIST, CIFAR-10, and CIFAR-100 datasets, respectively. Additionally, resource efficiency improvements are notable, reaching up to 27%, 34%, 24%, and 38% on the MNIST, Fashion-MNIST, CIFAR-10, and CIFAR-100 datasets, respectively. Moreover, the proposed HLIF neurone significantly improves temporal processing, adaptive inference capabilities, energy efficiency, and classification reliability, positioning it as a promising method for optimising deep spiking neural networks.