稳健损失函数的课程视角（英文版）

英文标题：A Curriculum View of Robust Loss Functions中文摘要：本文研究了鲁棒损失函数在标签噪声下的训练动态，通过重新表示大多数损失函数为具有相同类得分边界和不同样本加权函数的形式，提供了对训练动态的直观分析，证明了简单的修复可使欠拟合的鲁棒损失函数竞争力与最先进的方法相当，而训练时间表可以明显影响噪声鲁棒性，即使使用鲁棒损失函数。英文摘要：Robust loss functions are designed to combat the adverse impacts of labelnoise, whose robustness is typically supported by theoretical bounds agnosticto the training dynamics. However, these bounds may fail to characterize theempirical performance as it remains unclear why robust loss functions canunderfit.