Scholarlink

<p>Class imbalance remains a critical challenge in deep learning, often leading to biased models and degraded performance on underrepresented classes. This paper proposes Adaptive Confidence-Weighted Regularization (ACWR), a novel training framework that dynamically adjusts sample contributions based on prediction confidence during optimization. Unlike traditional data-level techniques such as synthetic oversampling, ACWR operates directly at the loss function level and prioritizes uncertain instances without generating additional data, resulting in improved generalization and reduced bias. We evaluate ACWR across a variety of standard image classification tasks and compare it with both conventional convolutional neural networks (CNNs) and advanced solutions including residual architectures, feature-level balancing methods, and ensemble-based approaches. Experimental results show that ACWR consistently outperforms baseline and state-of-the-art methods, achieving up to 38% relative accuracy improvement in highly imbalanced settings. Additionally, ACWR maintains competitive performance on more balanced datasets, confirming its robustness, efficiency, and adaptability to different data distributions and model 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

Adaptive confidence-weighted regularization for enhancing CNN performance on imbalanced datasets

  • 期刊论文

  • 发表日期：

    2025年11月17日

查看全文
  • Rasool Aliasgharpoor mooziraji,
  • Faraein Aeini,
  • Ebrahim Akbari,
  • Homayoun Motameni

摘要

Class imbalance remains a critical challenge in deep learning, often leading to biased models and degraded performance on underrepresented classes. This paper proposes Adaptive Confidence-Weighted Regularization (ACWR), a novel training framework that dynamically adjusts sample contributions based on prediction confidence during optimization. Unlike traditional data-level techniques such as synthetic oversampling, ACWR operates directly at the loss function level and prioritizes uncertain instances without generating additional data, resulting in improved generalization and reduced bias. We evaluate ACWR across a variety of standard image classification tasks and compare it with both conventional convolutional neural networks (CNNs) and advanced solutions including residual architectures, feature-level balancing methods, and ensemble-based approaches. Experimental results show that ACWR consistently outperforms baseline and state-of-the-art methods, achieving up to 38% relative accuracy improvement in highly imbalanced settings. Additionally, ACWR maintains competitive performance on more balanced datasets, confirming its robustness, efficiency, and adaptability to different data distributions and model architectures.