基于分段损失的生成对抗网络  

Generative Adversarial Network Based on Piecewise Loss

作　　者：刘其开 姜代红[2] 李文吉[3] LIU Qikai;JIANG Daihong;LI Wenji(School of Information and Control Engineering,China University of Mining and Technology,Xuzhou,Jiangsu 221116,China;School of Information and Electrical Engineering,Xuzhou University of Technology,Xuzhou,Jiangsu 221111,China;Key Laboratory of Airborne Geophysics and Remote Sensing Geology,Ministry of Land and Resources,China AerospaceGeophysical Survey and Remote Sensing Center for Land and Resources,Beijing 100083,China)

机构地区：[1]中国矿业大学信息与控制工程学院,江苏徐州221116 [2]徐州工程学院信电工程学院,江苏徐州221111 [3]中国国土资源航空物探遥感中心国土资源部航空地球物理与遥感地质重点实验室,北京100083

出　　处：《计算机工程》2019年第5期155-160,168,共7页Computer Engineering

基　　金：国家自然科学基金(51574232);国土资源部航空地球物理与遥感地质重点实验室航遥青年创新基金(2016YFL02);徐州市科技计划项目(KC16SQ78)

摘　　要：生成对抗网络(GAN)在训练过程中未能有效进行生成器与鉴别器间的同步更新,导致模型训练不稳定并出现模式崩溃的现象。为此,提出一种基于分段损失的生成对抗网络PL-GAN。生成器在不同的训练时期采用不同形式的损失函数,同时引入真实样本与生成样本之间的特征级损失,从而使鉴别器提取的特征更具有鲁棒性。MNIST和CIFAR-10数据集上的实验结果表明,与regular GAN、feature-wise GAN相比,PL-GAN具有更高的分类精度与运行效率。Generative Adversarial Network(GAN) fails to effectively execute the synchronous update between generator and discriminator during training,resulting in unstable model training and mode collapse.To solve this problem,a generative adversarial network PL-GAN based on piecewise loss is proposed.The generator uses different loss functions in different training periods,and introduces the feature-wise loss between the real sample and the generated sample,which makes the feature extracted by the discriminator more robust.Experimental results on MNIST and CIFAR-10 datasets show that PL-GAN has higher classification accuracy and operation efficiency than regular GAN and feature-wise GAN.

关 键 词：生成对抗网络 模式崩溃 特征级损失 分段损失 半监督学习 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]