BriefGPT.xyz
Nov, 2015
往深层网络添加梯度噪声可改善学习效果
Adding Gradient Noise Improves Learning for Very Deep Networks
HTML
PDF
Arvind Neelakantan, Luke Vilnis, Quoc V. Le, Ilya Sutskever, Lukasz Kaiser...
TL;DR
本文研究了在训练基本和复杂神经网络时添加梯度噪声的效果,发现这种技术可以显著提高训练效果,尤其是在较深的网络结构中更加有效,鼓励将该技术应用于更多复杂现代架构中。
Abstract
Deep feedforward and
recurrent networks
have achieved impressive results in many perception and language processing applications. This success is partially attributed to architectural innovations such as convolutional and long short-term
→