无监督网络对抗生成模型的研究-《计算机技术与发展》

文章信息/Info

作者:: 王晟全¹ ; 陈济颖²; 王奕翔³; 李昂¹; 4 ; 李鑫¹; 1. 南京理工大学紫金学院,江苏南京 210023; 2. 中国石油大学(华东) 储运与建筑工程学院,山东青岛 266580; 3. 广西大学,广西南宁 530003; 4. 南京邮电大学通信学院,江苏南京 210003

Author(s):: WANG Sheng-quan¹ ; CHEN Ji-ying² ; WANG Yi-xiang³ ; LI Ang¹; 4 ; LI Xin¹; 1. Nanjing University of Science and Technology Zijin College,Nanjing 210023,China; 2. College of Pipeline and Civil Engineering,China University of Petroleum,Qingdao 266580,China; 3. Guangxi University,Nanning 530003,China; 4.?School of Telecommunications &Information Engineering,Nanjing University of Posts and Telecommunications,Nanjing210003,China

Keywords:: unsupervised network; generative adversarial model; discriminant model; deep learning

摘要:: 一般的学习模型都是基于一个假设的随机分布,然后通过训练真实数据来拟合出模型。网络模型复杂并且数据集规模也不小,这种方法简直就是凭借天生蛮力解决问题。Goodfellow 认为正确使用数据的方式,是先对数据集的特征信息有 insight 之后,再干活。无监督学习是当下较为流行的话题,但也是困难较为繁多的话题。目前无监督学习可以分成以下两类,分别是确定型的自编码方法以及概率型的受限波尔兹曼机,其目标主要是使受限玻尔兹曼机达到稳定状态时原数据出现的概率最大。如何更快速更有效地地搭建模型以及如何做实验并有效地获得相关的实验结论是人们讨论的重点在研究中,在判别模型中增加正则化,用卷积层代替池化层,在生成模型中输出层使用 tanh 激活函数激活,这样使得最终运算的准确率和损失率大大下降,并减少了冗余成分。

Abstract:: The general learning model is based on a hypothetical random distribution and then fits the model by training the real data. The network model is complex with the small size of the data set,which is not conductive to solving the problem. Goodfellow believes that the correct way to use the data? ?is to have insight into the characteristics of the data set before doing the work. Unsupervised learning is apopular topic at the moment,but it is also? ? a topic with many difficulties. At present,unsupervised learning can be mainly divided into the following two categories:deterministic self-encoding method and probabilistic restricted Boltzmann machine. The probabilistic restricted Boltzmann machine aims to maximize the probability of occurre-? nce of the original data when reaching a steady state. How to build the model more quickly and efficiently and how to do experiments and obtain relevant experimental conclusions effectively are the focus of discussion. In this study,regularization is added in the discriminant model,the convolu- tion layer is used to replace the pooling layer,and the output layer in the generated model is activated by tanh activation function,which greatly reduces the accuracy and loss rate of the final operation as well as the redundant components.