Scholarlink

In this work, we propose a novel spiking neural network learning method that leverages correlations explicitly, thus providing a straightforward way of conveying the feedback signal to the network. The method implementation is based on converting an input vector into an array of spike sequences that are correlated with a given output sequence. The correlation is achieved by interpreting input feature values as Pearson’s correlation coefficients and randomly selecting spikes from the reference sequence. The attractiveness of this approach lies in facilitating the development of compact neural networks that do not rely on global error backpropagation training algorithms, which are difficult to implement on low-power-consuming analog hardware. The efficiency of the method is shown by experiments on several benchmark datasets: F1-macro of 99  ±  1% on Fisher’s Iris; 95  ±  2% on Wisconsin Breast Cancer; 92  ±  2% on Scikit-Learn Handwritten Digits. Additionally, we show that the proposed approach can also achieve competitive accuracies using memristive plasticity models, demonstrating the prospective possibility of its implementation in memristor-based neuromorphic hardware.

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

Direct Correlational Spike-Timing-Dependent Plasticity Learning Applied to Classification Tasks

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Alexander Sboev,
  • Dmitry Kunitsyn,
  • Yury Davydov,
  • Danila Vlasov,
  • Alexey Serenko,
  • Roman Rybka,
  • Yuanchao Liu

摘要

In this work, we propose a novel spiking neural network learning method that leverages correlations explicitly, thus providing a straightforward way of conveying the feedback signal to the network. The method implementation is based on converting an input vector into an array of spike sequences that are correlated with a given output sequence. The correlation is achieved by interpreting input feature values as Pearson’s correlation coefficients and randomly selecting spikes from the reference sequence. The attractiveness of this approach lies in facilitating the development of compact neural networks that do not rely on global error backpropagation training algorithms, which are difficult to implement on low-power-consuming analog hardware. The efficiency of the method is shown by experiments on several benchmark datasets: F1-macro of 99  ±  1% on Fisher’s Iris; 95  ±  2% on Wisconsin Breast Cancer; 92  ±  2% on Scikit-Learn Handwritten Digits. Additionally, we show that the proposed approach can also achieve competitive accuracies using memristive plasticity models, demonstrating the prospective possibility of its implementation in memristor-based neuromorphic hardware.