Ping Luo (羅平)
Ping Luo (羅平)
Home
News&Talks
Publications
Members
Projects
Contact
Light
Dark
Automatic
MetaLearning
Understanding Normalization in Deep Learning
一个卷积层,一个归一化层,一个非线性激活函数一起构成了深度卷积神经网络 (ConvNet)的“原子”结构。通过该基础结构的堆叠,产生了许多应用广泛的神经网络。归一化方法是这些神经网络的重要组成部分之一。本次报告的内容围绕深度学习的归一化方法展开,及其为神经网络带来的正则能力与泛化能力。报告分为四个部分。第一部以白化网络(Whitened Neural Network,WNN)为例,浅析神经网络前向计算与反向传播的关系:表明修改前向计算,将可以直接影响深度网络的费希尔信息矩阵 (Fisher …
Switchable Normalization
Meta-learning to learn normalization method for each hidden layer in ConvNet.
Learning-to-Learn-to-Normalize: Algorithms, Applications and Theory
Introduce a new family of normalization methods in Deep Learning.
Cite
×