Scholarlink

<p>Although deep neural networks have made significant progress in confidence calibration, conventional methods like temperature scaling and histogram binning face notable limitations in dynamic open-world scenarios due to their reliance on global static parameters or idealized data distribution assumptions. To address these challenges, we propose a Dynamic Confidence Propagation and Alternating Normalization (DCP-AN) framework. Our approach introduces three key innovations: (1) a bidirectional alternating propagation mechanism that enables sample-class confidence synergy through entropy-driven horizontal normalization and KL-divergence-weighted vertical normalization; (2) an adaptive temperature field with dynamic coefficients that achieves differential calibration for non-uniform confidence biases; and (3) a theoretically guaranteed spectral convergence process within 15 iterations. Extensive experiments demonstrate that DCP-AN achieves remarkable improvements: on ImageNet-LT, it boosts tail-class accuracy by 10.3% and reduces expected calibration error by 56%; in cross-domain adaptation, it decreases domain discrepancy by 24% while improving target domain accuracy by 5.5%. Furthermore, DCP-AN maintains computational efficiency with a GPU latency of 1.05&#xa0;ms and memory overhead under 0.5&#xa0;MB, making it suitable for real-time deployment.</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

Calibrating deep classifiers with dynamic confidence propagation and adaptive normalization

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月31日

查看全文
  • Peihua He,
  • Wei Fu,
  • Lufeng Wang,
  • Tian Fu

摘要

Although deep neural networks have made significant progress in confidence calibration, conventional methods like temperature scaling and histogram binning face notable limitations in dynamic open-world scenarios due to their reliance on global static parameters or idealized data distribution assumptions. To address these challenges, we propose a Dynamic Confidence Propagation and Alternating Normalization (DCP-AN) framework. Our approach introduces three key innovations: (1) a bidirectional alternating propagation mechanism that enables sample-class confidence synergy through entropy-driven horizontal normalization and KL-divergence-weighted vertical normalization; (2) an adaptive temperature field with dynamic coefficients that achieves differential calibration for non-uniform confidence biases; and (3) a theoretically guaranteed spectral convergence process within 15 iterations. Extensive experiments demonstrate that DCP-AN achieves remarkable improvements: on ImageNet-LT, it boosts tail-class accuracy by 10.3% and reduces expected calibration error by 56%; in cross-domain adaptation, it decreases domain discrepancy by 24% while improving target domain accuracy by 5.5%. Furthermore, DCP-AN maintains computational efficiency with a GPU latency of 1.05 ms and memory overhead under 0.5 MB, making it suitable for real-time deployment.