Disentangling Representation of Variational Autoencoders Based on Cloud Models

Abstract

Variational autoencoder (VAE) has the problem of uninterpretable data generation process, because the features contained in the VAE latent space are coupled with each other and no mapping from the latent space to the semantic space is established. However, most existing algorithms cannot understand the data distribution features in the latent space semantically. In this paper, we propose a cloud model-based method for disentangling semantic features in VAE latent space by adding support vector machines (SVM) to feature transformations of latent variables, and we propose to use the cloud model to measure the degree of disentangling of semantic features in the latent space. The experimental results on the CelebA dataset show that the method obtains a good disentangling effect of semantic features in the latent space, which proves the effectiveness of the method from both qualitative and quantitative aspects.