基于互信息约束的生成对抗网络分类模型

Classification models based on generative adversarial networks with mutual information regularization

传统的机器学习方法需要大量的含标注数据集来训练模型,并且容易引发过拟合,而生成对抗网络可以无监督地进行训练。此外,互信息约束能够让模型生成指定类别的数据,可用于扩充数据集。提出InfoCatGAN和C-InfoGAN两种模型,前者在CatGAN的基础上增加了互信息约束,使得生成的图片更加逼真;后者使用InfoGAN模型中的辅助网络Q做分类,能够在生成高质量图片的同时,达到较好的分类准确率。二者均能通过隐变量控制生成图片的类别,这对数据增强具有一定意义。另外,在加入少量标签信息之后,模型的准确率能有所提升。

This paper studies classification models based on generative adversarial networks with mutual information regularization.Traditional machine learning methods rely on a large number of labeled datasets,which are scarce in practice,to train the model and can easily overfit to spurious correlations in the data;while generating adversarial networks can be trained in an unsupervised manner.In addition,mutual information constraint allows the model to generate data of a specified category,which can be used to expand the data set.This paper proposes the InfoCatGAN and C-InfoGAN classification models.The former adds the mutual information term to CatGAN model in order to generate images of higher visual fidelity;the latter uses the InfoGAN model for classification,which can ensure the quality of the generated images and provide a mentionable classification accuracy.Additionally,both two models can control the category of generated images through latent variables,which has a certain significance for data augmentation.Moreover,after adding a small amount of label information,the accuracy of the model can be improved.