BriefGPT.xyz
May, 2019
深度神经网络初始化引起的泛化误差类型
A type of generalization error induced by initialization in deep neural networks
HTML
PDF
Yaoyu Zhang, Zhi-Qin John Xu, Tao Luo, Zheng Ma
TL;DR
通过利用DNN训练动力学在NTK领域中的线性性,本研究提供了关于DNN学习的初始化和损失函数如何影响其泛化误差的明确定量答案 ,证明了在NTK范围内通过抗对称初始化技巧(ASI)减少初始输出引起的误差的可能性并加速训练。
Abstract
How different initializations and loss functions affect the learning of a
deep neural network
(DNN), specifically its
generalization error
, is an important problem in practice. In this work, focusing on regressio
→