BriefGPT.xyz
Sep, 2019
XNOR-Net++: 二值神经网络的改进
XNOR-Net++: Improved Binary Neural Networks
HTML
PDF
Adrian Bulat, Georgios Tzimiropoulos
TL;DR
本文提出了一种改进的训练算法,用于针对二元神经网络的权重和激活二进制数的训练,提出了一种新的比XNOR-Net更优的方法,通过反向传播学习判别式融合激活和权重缩放因子。实验表明,与分析计算的结果相比,我们的方法更加准确,且在相同的计算预算下,可在 ImageNet 分类任务中提供高达 6% 的准确性提升。
Abstract
This paper proposes an improved training algorithm for
binary neural networks
in which both weights and activations are binary numbers. A key but fairly overlooked feature of the current state-of-the-art method of
xnor-
→