Research And Application Of Representation Learning Based On Variational Auto-encoder

In recent years,deep learning technology has achieved great success in computer vision,speech recognition,and machine translation.The research of Representation Learning plays an important role in the progress of deep learning.Learning representations of the data makes it easier to extract useful information when building classifiers or other predictors.Good representation of the data has great significance for improving the performance of deep learning algorithms.Deep learning technology now is widely used in many fields and has achieved the best performance.Researchers are more interested in supervised deep learning model as label information is so important to the model and can improve model performance.The rapid growth of data size make it more difficult to obtain label information of all data,semi-supervised and unsupervised learning methods become more and more important.In recent years,the research of the deep generative model has achieved widespread success.For example,Variational Auto-Encoder is one of the most popular unsupervised learning method for learning complex distributions.It builds on the neural network and can be trained with a stochastic gradient descent method.In this paper,we propose a new semi-supervised classification model within the framework of Variational Auto-Encoder.Specifically,multiple gaussian distribution are used in latent space for different class of data points.Training with a small amount of labeled data points can match different types of data to different gaussian distributions and can achieve good classification results.Experimental results show that our proposed model can achieve competitive classification performance with a small amount of labeled data.In this paper,we propose an unsupervised clustering method combining Gaussian Mixture EM algorithm and Variational Auto-Encoder.The method uses variational inference to approximate the posterior probability distribution as a gaussian mixture distribution,morevoer,it can be used for clustering,and the EM algorithm is used to estimate the clustering assignment in the hidden space.Experimental results show that our proposed method can achieve competitive performance.Moreover,by the method's generative nature,we show its capability of generating realistic samples for any specified cluster without using supervised information during training.