Research On Improved Algorithm Of Variational Autoencoder Based On Posterior Approximation

Accurate modeling of the complex generation process of high-dimensional data(such as images)is a key task of deep learning.In many application fields,Variational Autoencoders(VAE)have proven to be very effective for this task,and have the ability to interpret and directly control the latent variables corresponding to the latent hidden factors in the data generation.However,the traditional VAE also has shortcomings,that is,the performance of complex models is poor,and the generated images are often more blurry.This paper will improve the objective function of the model and change the structure of the model to solve the above problems.In this paper,a variational laplace autoencoders model based on determinantal point process(VLAE-DPP)is proposed.The method of determinantal point process is introduced into the variational laplace autoencoders model,and an unsupervised penalty loss is added on the basis of the original objective function to improve the quality of generated images.The VLAE-DPP model uses the Determinantal Point Process(DPP)to capture the similar diversity of real data,and then learns the kernel by extracting features from the encoder.Finally,the decoder is trained to optimize the loss between the pseudo,real,eigenvalues and eigenvectors of the kernel to encourage the decoder to simulate the diversity of real data,thereby generating high-quality images.This paper also changes the structure of the original variational laplace autoencoders,and inserts a discriminator behind the decoder,called the Adversarial Variational Laplace Autoencoders(AVLAE).The original variational laplace autoencoders uses the square error as the reconstruction target,which will affect the effect of the generated image to some extent.The AVLAE model proposed in this paper uses the feature representation of the discriminator to replace the original reconstruction error,and the discriminator and decoder have a process of adversarial learning,which can improve the quality of the generated image.The experimental results show that the VLAE-DPP model and AVLAE model can solve the problem of unclear image generation to varying degrees on multiple data sets,and improve the quality of generated images.