Research On Representation Learning Based On Variational Auto-encoder

With the development of machine learning,unsupervised representation learning has become one of the most important research directions.Variational autoencoder(VAE)have been regarded as one of the most valuable models in the field of unsupervised representation learning because of its explicit generative modeling approach.Unsupervised clustering representation learning is an important research branch in the field of representation learning.Traditional deep clustering representation learning methods pay more attention to extracting hidden layer features of data through deep neural networks to improve clustering accuracy,and less analyze the determinism of data categories in clustering tasks.At same time,there is a lack of analysis of the discrete latent vector distribution after imposing constraints.In this paper,we propose a variational deep generative clustering model based on entropy regularizations(VDGC-ER),which uses the variational auto-encoder as the basic framework and models a gussian mixture prior on its latent space.We first propose the sample entropy regularization term to the discrete latent vector of Gaussian mixture model to improve the clustering accuracy of the model.Further,we define the aggregated sample entropy regularization term on the discrete latent vector to reduce the clustering imbalance,so that it can avoid local optimization and improve the generative diversity.Then,we use the Monte Carlo sampling and re-parameterization trick to estimate the optimization objective of VDGC-ER model,and use the stochastic gradient descent method to calculate the model parameters.Finally,we design the comparison experiments on different datasets to demonstrate the performance of the VDGC-ER model.The experiment results show that the model can not only generate high-quality samples,but also present high-accuracy clustering.Unsupervised disentanglement representation learning is another important research branch in the field of representation learning.Traditional disentanglement representation learning models model the latent space of data in a relatively simple way,but in real-world datasets,the physical meanings of different data generative factors are often completely inconsistent,resulting in the limited capabilities of these disentanglement representation models,and not robust on the complex datasets.In this paper,we propose a Hierarchical Disentanglement Method based on VAE with Structured Priors(SP-HDVAE).The method first uses a Gaussian mixture model to model latent vector priors,learns structured latent space representations,and achieves coarse-grained disentangled representations.Then,the regular term inductive preference is introduced into the structured representation of the latent space,and the fine-grained disentanglement representation is realized on the basis of the structured representation,so that the SP-HDVAE model can realize the hierarchical decoupling representation as a whole.Finally,by designing comparative experiments on the d Sprites,the MNIST,the Fashion-MNIST,the 3dchair,the celeb A and the hybrid datasets,it is verified that the SP-HDVAE model has a good disentanglement ability and can be widely used to more complex datasets.