BriefGPT.xyz
May, 2018
添加一个神经元可以消除所有不良局部极小值
Adding One Neuron Can Eliminate All Bad Local Minima
HTML
PDF
Shiyu Liang, Ruoyu Sun, Jason D. Lee, R. Srikant
TL;DR
本研究旨在探讨神经网络中非凸损失函数的分析难点，通过添加具有跳跃连接的特殊神经元或每层添加一个特殊神经元，证明了在温和的条件下，每个局部最小值都是全局最小值。
Abstract
One of the main difficulties in analyzing
neural networks
is the non-convexity of the
loss function
which may have many bad
local minima
.
→