标签平滑，作者说one-hot这种脉冲式的标签导致过拟合

new_labels = (1.0 - label_smoothing) * one_hot_labels + label_smoothing / num_classes

Szegedy在网络实现的时候，令 label_smoothing = 0.1，num_classes = 1000。Label smooth提高了网络精度0.2%