BriefGPT.xyz
Jun, 2020
深度学习中的方向收敛和对齐
Directional convergence and alignment in deep learning
HTML
PDF
Ziwei Ji, Matus Telgarsky
TL;DR
本文证明了通过梯度流学习方法得到的深层同质网络权重会趋向于收敛,并阐述了相应的研究内容,包括但不限于梯度流、分类损失、边缘最大化、显著图等方面。
Abstract
In this paper, we show that although the minimizers of cross-entropy and related
classification losses
are off at infinity, network weights learned by
gradient flow
converge in direction, with an immediate coroll
→