New Focal Calibration Loss aims to improve classifier probability accuracy

Researchers have introduced Focal Calibration Loss (FCL), a novel objective for deep neural classifiers designed to improve confidence calibration in probability outputs. FCL integrates a squared probability-error term into the focal loss, aiming to control posterior distortion. While FCL is classification-calibrated, it is not always a proper scoring rule, meaning its optimal probability vector may deviate from the true posterior. The study quantifies this displacement and demonstrates that the FCL anchor parameter ($\lambda$) effectively manages it, offering theoretical guarantees on calibration error. Experiments on various datasets like CIFAR-10/100 and Tiny-ImageNet show FCL is competitive, though performance varies by metric and task, with AdaFocal sometimes outperforming it in binned calibration error. AI

影响 Introduces a new method for improving the reliability of probability estimates from deep learning models, potentially impacting applications where confidence scores are critical.

排序理由 Academic paper detailing a new loss function for deep neural classifiers. [lever_c_demoted from research: ic=1 ai=1.0]

在 arXiv stat.ML 阅读 →

• AdaFocal
• arXiv
• Brier
• CIFAR-10
• CIFAR-100
• Focal Calibration Loss
• Tiny-ImageNet

• 论文
• 模型发布

AI 生成摘要 · Google Gemini · 来自 1 个来源。 我们如何撰写摘要 →