Scholarlink

Bias in race and gender classification models creates strong ethical dilemmas, where disparities typically exist among groups of different demographics. In this research, a Bias-Aware Grad-CAM method is proposed that embeds interpretability-driven bias reduction within training. A ResNet model was trained for gender and race classification using the UTKFace dataset, which shows preliminary errors in minority group predictions. The Bias-Aware Grad-CAM framework proposed produces heatmaps to detect model attention areas, calculating Intersection over Union (IoU) between the heatmaps and pre-defined face bounding boxes. A reweighting process in the loss function, based on IoU scores, is introduced to retrain the model. Results show enhanced classification performance by aligning model attention with appropriate facial areas better. This strategy emphasizes the promise of explainability-guided methods to real-time bias reduction in face recognition models.

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

Exploring Biases and Interpretability of Deep Learning Models

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Shreya Pulluri,
  • Megha Mohan,
  • Madduri Aditya Vardhan Reddy,
  • Chitresh Simhadri,
  • M. Geetha

摘要

Bias in race and gender classification models creates strong ethical dilemmas, where disparities typically exist among groups of different demographics. In this research, a Bias-Aware Grad-CAM method is proposed that embeds interpretability-driven bias reduction within training. A ResNet model was trained for gender and race classification using the UTKFace dataset, which shows preliminary errors in minority group predictions. The Bias-Aware Grad-CAM framework proposed produces heatmaps to detect model attention areas, calculating Intersection over Union (IoU) between the heatmaps and pre-defined face bounding boxes. A reweighting process in the loss function, based on IoU scores, is introduced to retrain the model. Results show enhanced classification performance by aligning model attention with appropriate facial areas better. This strategy emphasizes the promise of explainability-guided methods to real-time bias reduction in face recognition models.