Scholarlink

<p>Biological neural systems adapt and reorganise with a flexibility that artificial models still struggle to compete with. Building on this idea, we introduce a Neuroplastic Neural Network (NPNN) framework that explores how key aspects of neuroplasticity might improve artificial learning. The model brings together four biologically inspired mechanisms: connection pruning, Hebbian learning, connection requalification and adaptive learning rate, within a conventional feedforward structure trained by backpropagation. The NPNN was evaluated on four benchmark datasets (MNIST, FashionMNIST, CIFAR-10, and CIFAR-100) across ten independent training runs, and a series of ablation studies to isolate each mechanisms’ individual contributions. Across all datasets, the NPNN outperformed a standard fully connected model, with statistically significant improvements, particularly on complex CIFAR tasks. On average, validation accuracy rose by around 9% on CIFAR-10 and 21% on CIFAR-100 (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(p \le 0.001\)</EquationSource> </InlineEquation>), alongside 14% faster accuracy convergence and a roughly one-third reduction in loss. Ablation studies show that Hebbian learning and adaptive learning rates play the largest roles, while their combination with structural plasticity produces the highest overall gains. An extended CNN study further confirmed these findings, with the Neuroplastic CNN achieving consistent improvements of between 1.5 to 5% in accuracy, recall, and F1 score over a standard CNN baseline on both CIFAR datasets. These results suggest a promising direction for biologically informed learning architectures.</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

Neuroplastic Neural Networks: Adaptive Learning Through Structural Plasticity and Hebbian Updates

  • 期刊论文

  • 发表日期：

    2026年5月26日

查看全文
  • Abrar Fahyaz,
  • Kaveh Kiani,
  • Maybin K. Muyeba

摘要

Biological neural systems adapt and reorganise with a flexibility that artificial models still struggle to compete with. Building on this idea, we introduce a Neuroplastic Neural Network (NPNN) framework that explores how key aspects of neuroplasticity might improve artificial learning. The model brings together four biologically inspired mechanisms: connection pruning, Hebbian learning, connection requalification and adaptive learning rate, within a conventional feedforward structure trained by backpropagation. The NPNN was evaluated on four benchmark datasets (MNIST, FashionMNIST, CIFAR-10, and CIFAR-100) across ten independent training runs, and a series of ablation studies to isolate each mechanisms’ individual contributions. Across all datasets, the NPNN outperformed a standard fully connected model, with statistically significant improvements, particularly on complex CIFAR tasks. On average, validation accuracy rose by around 9% on CIFAR-10 and 21% on CIFAR-100 ( \(p \le 0.001\) ), alongside 14% faster accuracy convergence and a roughly one-third reduction in loss. Ablation studies show that Hebbian learning and adaptive learning rates play the largest roles, while their combination with structural plasticity produces the highest overall gains. An extended CNN study further confirmed these findings, with the Neuroplastic CNN achieving consistent improvements of between 1.5 to 5% in accuracy, recall, and F1 score over a standard CNN baseline on both CIFAR datasets. These results suggest a promising direction for biologically informed learning architectures.