长尾数据分类面临多数类主导问题,解耦训练范式通过特征学习与分类器再训练分离取得进展。创新指标Logits Magnitude和Regularized Standard Deviation揭示了分类器优化的关键。提出的标签过度平滑(LOS)方法软化类间差异,有效抑制多数类干扰,在多个长尾数据集上实现最佳性能。
Had I not seen the Sun
------- by Emily Dickinson
Had I not seen the Sun
I could have borne the shade
But Light a newer Wilderness
My Wilderness has made —