Research On Construction Method Of Semi-supervised Deep Generation Model

With the advent of the era of big data,people’s demand for data is increasing,and traditional supervised learning has been unable to meet people’s needs.Semi-supervised learning has become a hot research topic because it only requires a small number of labeled samples.At the same time,with the continuous development of deep learning,the application of deep learning in semi-supervised learning has become more and more extensive and diversified.Among them,the semi-supervised deep generative model combining deep generative model and semi-supervised learning has become one of the current mainstream ideas.The commonly used construction methods of semi-supervised depth generation model are the approximate method(auto-encoder)to obtain the likelihood function by variational or sampling method and the implicit method(generative adversarial networks)to avoid obtaining the likelihood function.However,there are still some problems in the two construction methods,such as fuzzy sample generation,insufficient classification performance,and model collapse.To solve the current problems of the semi-supervised depth generation model,this paper proposes the following two improved methods for constructing the semi-supervised depth generation model.The specific research work and innovations are as follows.(1)Aiming at the problem that the semi-supervised generation model based on autoencoder ignores the information contained in the unlabeled data in the sample,and the semisupervised classifier does not learn the overall distribution of the original data,a semisupervised depth generation model based on Wasserstein distance(WCVAE)is constructed.The Wasserstein auto-encoder(WAE)is improved to solve the semi-supervised learning problem.By using the improved generation model,the semi-supervised depth generation model considers the marginal distribution of all samples and the conditional distribution of labeled samples,and improves the quality of the semi-supervised depth generation model.At the same time,the adversarial structure in hidden variable space in WAE model is further optimized,so that the marginal distribution of feature space between labeled samples and unlabeled samples is considered.The experimental results show that the proposed model reduces the error rate by1.8% compared with the current model,which can effectively improve the semi-supervised classification ability of the model.(2)Aiming at the problem of poor sample quality of semi-supervised generation model based on auto-encoder,a semi-supervised deep generation model based on Gaussian mixture model(GMA-WCVAE)is constructed.This model replaces the target distribution in the feature space from the original normal distribution to the mixed Gaussian distribution,and uses the Cramer-Wold distance as the measurement standard,so that the GMA-WCVAE model can learn the joint probability distribution of data and its categories,and further improve the quality of generated samples.The experimental results show that the FIDs value of GMA-WCVAE is 17.2%lower than that of the WCVAE model,which can effectively improve the quality of samples generated by the model.