yuyinzhou / L2B

This repository includes the official project of L2B, from our paper "Learning to Bootstrap for Combating Label Noise".
28 stars 3 forks source link

关于用梯度来reweight loss的困惑 #2

Closed Alex20172023 closed 2 years ago

Alex20172023 commented 2 years ago

作者您好!完美工作。 其实有一点不明白,为什么会用到loss在meta data上计算的梯度,来直接作为新的α和β呢? 我的意思是,这种参数的更新个人感觉应该是用α'=α-lr*梯度,直接用梯度作为更新后的值,有点不太理解具体什么意义。 希望得到作者的答复!

yuyinzhou commented 2 years ago

感谢关注与支持! 这部分主要是依据于 Ren, Mengye, et al. "Learning to reweight examples for robust deep learning." International conference on machine learning. 2018. "直接用梯度作为更新后的值"其实是经过一步online approximation之后的一个近似。具体推导请参考paper的第三章节。 再次感谢您的支持!