Scholarlink

<p>The quantization of deep neural networks (DNNs) is a vital technique aimed at minimizing the computational and memory demands of networks, particularly for their use in resource limited environments such as mobile devices and embedded systems. By employing lower bit widths for weights, activations, and gradients, quantization methods effectively decrease architecture size and enhance inference speed without a substantial loss in performance. Nevertheless, finding the right balance between architecture size and accuracy presents ongoing challenges, as more aggressive quantization methods can result in notable accuracy losses. In this paper, we introduce a variation of dynamic quantization. The proposed approach assigns non-uniform bit widths to each layer of the CNN (convolutional neural network) during the quantization process. Our novel framework integrates dynamic quantization with interpretability techniques typically utilized in traditional floating-point CNNs. This method leverages feature maps derived from floating-point CNNs to discern the intrinsic relationships among different network components. Specifically, we calculate gradient-based importance scores for each CNN layer, which inform the dynamic distribution of varying bit widths across the architecture. Moreover, we present two robust loss functions specifically designed for the quantization of CNNs during the QAT (quantization-aware training) phase. Extensive evaluations conducted on publicly accessible datasets, including CIFAR-10, CIFAR-100, ImageNet and AgeDB, reveal that our quantization technique achieves only a minimal accuracy drop of <b>0.7%</b> compared to the floating-point architectures, while also delivering a substantial reduction in computational requirements by <b>88.1%</b>. This makes our method particularly suitable for real-world application deployments.</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

SGDNI: Sensitivity Guided Dynamic Noise Injection for Robust Quantization of Deep Neural Networks

  • 期刊论文

  • 发表日期：

    2026年5月22日

查看全文
  • Aratrik Chattopadhyay,
  • Anamika Jha,
  • Mrinal Banerji,
  • Pushkar Nath,
  • Disha Jain

摘要

The quantization of deep neural networks (DNNs) is a vital technique aimed at minimizing the computational and memory demands of networks, particularly for their use in resource limited environments such as mobile devices and embedded systems. By employing lower bit widths for weights, activations, and gradients, quantization methods effectively decrease architecture size and enhance inference speed without a substantial loss in performance. Nevertheless, finding the right balance between architecture size and accuracy presents ongoing challenges, as more aggressive quantization methods can result in notable accuracy losses. In this paper, we introduce a variation of dynamic quantization. The proposed approach assigns non-uniform bit widths to each layer of the CNN (convolutional neural network) during the quantization process. Our novel framework integrates dynamic quantization with interpretability techniques typically utilized in traditional floating-point CNNs. This method leverages feature maps derived from floating-point CNNs to discern the intrinsic relationships among different network components. Specifically, we calculate gradient-based importance scores for each CNN layer, which inform the dynamic distribution of varying bit widths across the architecture. Moreover, we present two robust loss functions specifically designed for the quantization of CNNs during the QAT (quantization-aware training) phase. Extensive evaluations conducted on publicly accessible datasets, including CIFAR-10, CIFAR-100, ImageNet and AgeDB, reveal that our quantization technique achieves only a minimal accuracy drop of 0.7% compared to the floating-point architectures, while also delivering a substantial reduction in computational requirements by 88.1%. This makes our method particularly suitable for real-world application deployments.