Scholarlink

<p>Convolutional neural networks (CNNs) have achieved strong performance across various vision tasks, but their training remains sensitive to weight initialization. Traditional methods offer reasonable convergence but often suffer from vanishing gradients, poor variance control, or limited adaptability in deeper architectures. This paper introduces quantum-inspired optimal transport-based weight initialization (QIOT-WI), a method that combines uniform variance-scaled initialization with quantum-inspired perturbation and optimal transport mapping to improve weight distribution. The proposed approach aims to stabilize gradient flow and accelerate convergence without altering network design. Experiments on benchmark datasets across diverse vision tasks confirm that QIOT-WI provides consistent improvements in convergence stability and classification performance over conventional methods.</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

QIOT-WI: quantum-inspired weight initialization for convolutional neural networks

  • 期刊论文

  • 发表日期：

    2025年6月8日

查看全文
  • Abdelaadim Khriss,
  • Aissa Kerkour Elmiad,
  • Mohammed Badaoui

摘要

Convolutional neural networks (CNNs) have achieved strong performance across various vision tasks, but their training remains sensitive to weight initialization. Traditional methods offer reasonable convergence but often suffer from vanishing gradients, poor variance control, or limited adaptability in deeper architectures. This paper introduces quantum-inspired optimal transport-based weight initialization (QIOT-WI), a method that combines uniform variance-scaled initialization with quantum-inspired perturbation and optimal transport mapping to improve weight distribution. The proposed approach aims to stabilize gradient flow and accelerate convergence without altering network design. Experiments on benchmark datasets across diverse vision tasks confirm that QIOT-WI provides consistent improvements in convergence stability and classification performance over conventional methods.