今天大概了解雅克比归一化损失函数的思想,是对输出out_y= (64, 10), 输入x = (64, 28*28), 求 dy/dx,做为损失函数的一部分,进行梯度更新.
该损失的作用可以提高模型的扰动稳定性.
算法:
参考文献
Robust Learning with Jacobian Regularization
https://github.com/facebookresearch/jacobian_regularizer
下面这篇论文提出Spectral Norm Regularization
Spectral Norm Regularization for Improving the Generalizability of Deep Learning
思考:
1.在目标检测中,在类别损失计算的时候,加入类别损失中