(1)目标函數中加入正則化
(2)dropout
(3)batch normalization
(4)Label Smoothing Regularization(LSR)
LSR 就是為了緩解由label不夠soft而容易導緻過拟合的問題,使模型對預測less confident。
https://baijiahao.baidu.com/s?id=1572001686264680&wfr=spider&for=pc
https://blog.csdn.net/qiu931110/article/details/86684241
https://blog.csdn.net/edogawachia/article/details/78552257