Scholarlink

Abstract <p>In this work, we demonstrate the capability of the developed compact single-layer SNN with probabilistic spiking neurons trained using local synaptic plasticity and correlation-based algorithm to solve classification tasks. Three synaptic plasticity models are compared: traditional spike-timing-dependent plasticity (STDP) and approximations of conductance plasticity measured in nanocomposite and poly-p-xylylene memristors. Learning is governed by a simple local reinforcement criterion based on pre- and post-synaptic spike correlations. Despite its simplicity, the network achieves competitive performance on standard benchmarks—F1-scores of approximately 97<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\%\)</EquationSource> <!--BPhysMGU2570300Kunitsyn-m1--> </InlineEquation> on the Iris dataset, approximately 90<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\%\)</EquationSource> <!--BPhysMGU2570300Kunitsyn-m2--> </InlineEquation> on Breast Cancer Wisconsin, and approximately 88<InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(\%\)</EquationSource> <!--BPhysMGU2570300Kunitsyn-m3--> </InlineEquation> on Scikit-Learn Digits—using both conventional STDP and memristive plasticities. The resulting accuracies demonstrate that compact SNNs with local learning may be suited for the implementation on practical, low-power neuromorphic systems.</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

Probabilistic Spiking Neural Network with Correlation-Based Memristive Synaptic Update

  • 期刊论文

  • 发表日期：

    2025年12月1日

查看全文
  • D. Kunitsyn,
  • A. Sboev,
  • Y. Davydov,
  • D. Vlasov,
  • A. Serenko,
  • R. Rybka

摘要

Abstract

In this work, we demonstrate the capability of the developed compact single-layer SNN with probabilistic spiking neurons trained using local synaptic plasticity and correlation-based algorithm to solve classification tasks. Three synaptic plasticity models are compared: traditional spike-timing-dependent plasticity (STDP) and approximations of conductance plasticity measured in nanocomposite and poly-p-xylylene memristors. Learning is governed by a simple local reinforcement criterion based on pre- and post-synaptic spike correlations. Despite its simplicity, the network achieves competitive performance on standard benchmarks—F1-scores of approximately 97 \(\%\) on the Iris dataset, approximately 90 \(\%\) on Breast Cancer Wisconsin, and approximately 88 \(\%\) on Scikit-Learn Digits—using both conventional STDP and memristive plasticities. The resulting accuracies demonstrate that compact SNNs with local learning may be suited for the implementation on practical, low-power neuromorphic systems.