Scholarlink

<p>Contrastive techniques are widely employed in computer vision to reduce overfitting and enhance the generalization power of deep neural networks. However, existing techniques, such as contrastive learning, often suffer from longer training time, higher computational complexity, and suboptimal performance. These inefficiencies become more pronounced when training on large-scale datasets, underscoring the need for more efficient training methodologies. To address these issues, we propose a novel memory-based contrastive regularization technique that significantly enhances both training efficiency and model performance. The proposed model organizes diverse, non-overlapping input samples into bags, which are processed by the network to extract deep features. These features are then passed to a Memory Access Module (MAM), which searches for relevant items in memory and computes positive loss to introduce intra-class compactness using Euclidean similarity measures. Additionally, we introduce a negative center loss (NCL) regularizer that identifies the center points of each class, calculates the distances between them, and enforces inter-class separability through a negative loss term. Together, these properties of intra-class compactness and inter-class separability improve feature representation, enhance the model’s learning capability, and mitigate overfitting, ultimately leading to superior generalization. Our approach has been rigorously evaluated on multiple deep architectures using the CIFAR-10 and CIFAR-100 datasets, demonstrating consistent performance gains while reducing training time and computational complexity. On CIFAR-100 with the ResNet-18 architecture, our model achieves 83.40% classification accuracy, improving the cross-entropy baseline (81.30%) by 2.10%.</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

Memory-based batch contrastive regularization for enhanced feature learning in deep neural networks

  • 期刊论文

  • 发表日期：

    2026年3月7日

查看全文
  • Muhammad Tanveer,
  • Hong Yee Alvin Wong,
  • Sheikh Faisal Rashid,
  • Jun Luo

摘要

Contrastive techniques are widely employed in computer vision to reduce overfitting and enhance the generalization power of deep neural networks. However, existing techniques, such as contrastive learning, often suffer from longer training time, higher computational complexity, and suboptimal performance. These inefficiencies become more pronounced when training on large-scale datasets, underscoring the need for more efficient training methodologies. To address these issues, we propose a novel memory-based contrastive regularization technique that significantly enhances both training efficiency and model performance. The proposed model organizes diverse, non-overlapping input samples into bags, which are processed by the network to extract deep features. These features are then passed to a Memory Access Module (MAM), which searches for relevant items in memory and computes positive loss to introduce intra-class compactness using Euclidean similarity measures. Additionally, we introduce a negative center loss (NCL) regularizer that identifies the center points of each class, calculates the distances between them, and enforces inter-class separability through a negative loss term. Together, these properties of intra-class compactness and inter-class separability improve feature representation, enhance the model’s learning capability, and mitigate overfitting, ultimately leading to superior generalization. Our approach has been rigorously evaluated on multiple deep architectures using the CIFAR-10 and CIFAR-100 datasets, demonstrating consistent performance gains while reducing training time and computational complexity. On CIFAR-100 with the ResNet-18 architecture, our model achieves 83.40% classification accuracy, improving the cross-entropy baseline (81.30%) by 2.10%.