Research And Application Of Deep Probabilistic Statistics Generative Model

Because of its flexiblity of statistical inferance,the probablistic inference has been attracting a growing body of research,and deep learning has been widely used by virtue of its performance.So,how to combine them has been the direction to explore.This article marry ideas from deep neural networks and approximate probablistic inference to derive a generalized class of deep,directed generative models,endowed with a new algorithm for scalable inference and learning.By demonstrations on several applacations,such as data simulation,data visualization and image super-resolution,we obtain a model that is able to generate realistic samples of data,provides a powerfull tool for high-dimensional data visualisation,and performs well on the image super-resolution.The main contents of each section are summarized as follows:In the first part,this thesis presents some probablistic inference methods and some neural networks used in deep learning.Firstly,we introduce three kinds of probablistic inference methods and compare their performance on the mixed Gaussian model.Then we introduce several common neural networks and provide some optimization skills.Finally,we sum up the similarities and differences between neural networks and probablistic model.In the second part,we propose a deep,directed generative models – varational autoencoder(VAE),which is combined ideas from deep neural networks and probablistic inferance.Firstly,we describe the principle of VAE model.Then,we show that a reparameterization of the variational lower bound yields a lower bound estimator that can be straightforwardly optimized using standard stochastic gradient methods.Finally,we sum up the advantages of VAE model over autoencoder(AE)model.In the last part,we propose three applications based on the VAE model.The first application is data simulation for a variety of data,and we find that the VAE model has a high degree of superiority in data simulation.The second application is data visualization,and it is proved to be a powerful tool for high-dimensional data visualisation.The last application is image super-resolution,and our method performs well on the image super-resolution.