Regularization - Label Smoothing

Related to: Deep Learning

Label Smoothing

레이블 스무딩이란 Szegedy et al. (2016)이 제안한 기법으로, 라벨을 스무딩하여 모델 일반화 성능을 꾀합니다.

$y_k^{L.S.}=y_k(1-\alpha )+\alpha /K$

예시)

$y_1=[0,1,0,0]\alpha =0.1$

$$\begin{gathered} y_1^{L.S.}=y_1(1-0.1)+0.1/4 \\ =[0.25,0.925,0.025,0.025] \end{gathered}$$

참조

https://ratsgo.github.io/insight-notes/docs/interpretable/smoothing