GBCNs：用于提升1-bit DCNNs性能的基因二进制卷积网络
摘要
训练1-bit深度卷积神经网络（DCNNs）是计算机视觉中最具挑战性的问题之一，因为相较于常规DCNNs，它更容易陷入局部极小。原因在于1-bit DCNNs的二值化核与激活函数会导致显著的精度损失与训练低效。为解决该问题，我们提出基因二进制卷积网络（GBCNs）来优化1-bit DCNNs，通过引入一种新的平衡遗传算法（BGA）在端到端框架中提升表征能力。所提出的BGA方法用于修正GBCNs的二值化过程以缓解局部极小问题，可显著提升1-bit DCNNs的性能。我们开发了一种通用且灵活的新型BGA模块，可轻松集成到现有DCNNs中，如WideResNets和ResNets。在对象分类任务（CIFAR、ImageNet）上的大量实验验证了所提方法的有效性。值得强调的是，我们的方法在对象识别任务（即人脸识别、人脸与行人重识别）上表现出很强的泛化能力。
引用
@article{arxiv.1911.11634,
title = {GBCNs: Genetic Binary Convolutional Networks for Enhancing the Performance of 1-bit DCNNs},
author = {Chunlei Liu and Wenrui Ding and Yuan Hu and Baochang Zhang and Jianzhuang Liu and Guodong Guo},
journal= {arXiv preprint arXiv:1911.11634},
year = {2020}
}
备注
We withdraw it for error in Fig. 4. Yuan Hu, who is supervised by Chunlei Liu, forgot to set "tensorboard" for exp. Chunlei knew this and they thought ok regress several recorded real points on old data for Fig. 4. Max. accuracy is right. But made error due to insufficient communication. Hu and Liu are responsible for it. Other authors were not told details. We will correct code github.com/liuchunlei0430