Batch Normalization

 每個batch中的元素單位大小相同，有點像歸一化 優點： because of less covariate shift, learning rate可以設大一點 less vanishing  gradient problems less sensitive to initialization

>>阅读原文<<