一种面向神经网络谱归一化间隔界的 PAC-Bayes 方法
机器学习
2018-02-27 v2
摘要
我们以前馈神经网络层谱范数与权重 Frobenius 范数之积的形式，给出了一个泛化界。该泛化界通过 PAC-Bayes 分析推导得出。
引用
@article{arxiv.1707.09564,
title = {A PAC-Bayesian Approach to Spectrally-Normalized Margin Bounds for Neural Networks},
author = {Behnam Neyshabur and Srinadh Bhojanapalli and Nathan Srebro},
journal= {arXiv preprint arXiv:1707.09564},
year = {2018}
}
备注
Accepted to ICLR 2018