Batch Normalization

 每个batch中的元素单位大小相同，有点像归一化 优点： because of less covariate shift, learning rate可以设大一点 less vanishing  gradient problems less sensitive to initialization

>>阅读原文<<