摘要：The generative model, which can generate samples randomly by learning the probability density of observable data, has been widely concerned for the past few years. It has been successfully applied in a wide range of fields, such as image generation, image restoration, density estimation, natural language and speech recognition, style transfer and super resolution, and so on. Deep generative model with multiple hidden layers in the network structure becomes a research hotspot because of its better generation ability. Depending on the different methods of calculating the maximum likelihood function, we divide the models into three types: the first kind of method is the approximate method, which use the sampling method to calculate approximately the likelihood function, such as Restricted Boltzmann machines (RBM) and Deep belief network (DBN), Deep Boltzmann machines (DBM), helmholtz machine based on RBM. The alternatives are to optimize directly the variational lower bound of likelihood function, it is named as variational autoencoder. The important improvements to these variants include importance weighted autoencoders and auxiliary deep generative models; the second kind is implicit methods, the representative model is generative adversarial nets (GAN), GAN＇s model parameters is optimized by the adversaring behavior between the generator and the discriminator. The principal instantiations of GAN include Wasserstein GAN, deep convolutional generative adversarial networks and BigGAN. The third kind involve flow and neural autoregressive net, the main variations of the flow paradigm include normalizing flow based on nonlinear independent components estimation, invertible residual networks and variational inference with flow. The successful improvements to the neural autoregressive net include neural autoregressive distribution estimation, pixel recurrent neural network, masked autoencoder for distribution estimation and WaveNet. We outline the principle and structure of these deep generative models, and look forward to the future work. Copyright ?2022 Acta Automatica Sinica. All rights reserved.

ISSN号：0254-4156

卷、期、页：卷48期1：40-74

发表日期：2022-01-01

期刊分区（SCI为中科院分区）：无

收录情况：EI（工程索引）

发表期刊名称：Zidonghua Xuebao/Acta Automatica Sinica

通讯作者：胡铭菲

第一作者：左信,刘建伟

论文类型：期刊论文

论文概要：胡铭菲,左信,刘建伟，Survey on Deep Generative Model，Zidonghua Xuebao/Acta Automatica Sinica，2022，卷48期1：40-74

论文题目：Survey on Deep Generative Model