Scholarlink

<p>The loss function is the most important guiding factor in optimizing the model parameters when training a Convolutional Neural Networks (CNN). This effect is even more significant for imbalanced datasets, as it can bias the model to classify only the majority class labels. To address this, Cost-sensitive weighting is used, where the weight factor is multiplied by the loss value, allowing the model to focus more on the minority class samples. In this study, we propose a Dynamic Focal Loss that updates the loss contribution dynamically based on the dataset distribution and the predicted probability of input samples. This allows the model to focus more on the hard minority samples by increasing their weight contribution to the overall loss. This approach can be applied directly to various computer vision applications. The experiments on the benchmark datasets from Canadian Institute for Advanced Research (CIFAR) (i.e., CIFAR10 and CIFAR100) yield comparable results in terms of average <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="42979_2025_4198_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="18" /> </InlineMediaObject> <EquationSource Format="TEX">\(F_1\)</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mi>F</mi> <mn>1</mn> </msub> </math></EquationSource> </InlineEquation> scores.</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

Dynamic Focal Loss for Imbalanced Learning

  • 期刊论文

  • 发表日期：

    2025年7月16日

查看全文
  • Hafiz Mughees Ahmad,
  • Afshin Rahimi

摘要

The loss function is the most important guiding factor in optimizing the model parameters when training a Convolutional Neural Networks (CNN). This effect is even more significant for imbalanced datasets, as it can bias the model to classify only the majority class labels. To address this, Cost-sensitive weighting is used, where the weight factor is multiplied by the loss value, allowing the model to focus more on the minority class samples. In this study, we propose a Dynamic Focal Loss that updates the loss contribution dynamically based on the dataset distribution and the predicted probability of input samples. This allows the model to focus more on the hard minority samples by increasing their weight contribution to the overall loss. This approach can be applied directly to various computer vision applications. The experiments on the benchmark datasets from Canadian Institute for Advanced Research (CIFAR) (i.e., CIFAR10 and CIFAR100) yield comparable results in terms of average \(F_1\) F 1 scores.