摘要
由于反向传播算法、批正则化等方法的提出,基于深度学习的判别式模型在广泛应用中取得极大的成功。传统的生成式模型大部分基于最大化似然估计,对概率密度的估计十分困难,直接导致传统生成式模型难以计算和优化。生成对抗网络(Generative Adversarial Networks,GAN)可有效解决传统生成模型在概率密度估计方面的问题,实现良好的生成效果。基于此,综述GAN的基本原理、模型结构、应用领域及存在的问题,分析相关研究,并提出未来的研究方向。
Due to the proposal of methods such as backpropagation algorithm and batch regularization,discriminative models based on deep learning have achieved great success in wide applications.Most traditional generative models are based on maximum likelihood estimation,and the estimation of probability density is very difficult,which directly leads to the difficulty in calculating and optimizing traditional generative models.Generative Adversarial Networks(GAN)effectively solve the problems of traditional generative models in probability density estimation and achieve good generative effects.Based on this,the basic principles,model structure,application fields and existing problems of GAN are reviewed,the related research is analyzed,and the future research directions are proposed.
作者
黄桂振
梁婷
郑梦
魏嘉银
HUANG Guizhen;LIANG Ting;ZHENG Meng;WEI Jiayin(School of Data Science and Information Engineering,Guizhou Minzu University,Guiyang 550025,China)
出处
《智能物联技术》
2025年第4期17-20,共4页
Technology of Io T& AI
基金
贵州省科技计划项目(黔科合基础[2018]1082,黔科合基础[2019]1159,黔科合基础ZK[2021]一般300,黔科合基础ZK[2022]一般195,黔科合基础ZK[2023]一般143)
贵州省青年科技人才成长项目(黔教合KY字[2021]104)
贵州省教育厅自然科学研究项目(黔教技[2022]015号,黔教技[2022]047号,黔教技[2023]012号,黔教技[2023]061号,黔教技[2023]062号)
贵州民族大学基金科研项目(GZMUZK[2023]YB13)。
关键词
生成对抗网络(GAN)
生成器
判别器
无监督算法
Generative Adversarial Networks(GAN)
generators
discriminators
unsupervised algorithms
