VampPrior VAE with Coded Hierarchical Inputs in Quotient Space

Generative models for image generation and Large Language Models have attracted attention in recent years. The variational autoencoder (VAE) is a basic and standard generative model consisting of an encoder and decoder. β-VAE, which introduces a parameter β to balance the reconstruction and the data distribution approximation accuracy, enables data structure analysis by adjusting the parameter β. However, a trade-of generally exists between high-precision data generation, data distribution approximation, and data structure analysis, and achieving these requirements simultaneously is difficult. This paper proposes a new hierarchical VAE based on a variational mixture of posteriors prior VAE (VampPrior VAE) that uses a mixture of normal distributions to approximate the data distribution. This hierarchical structure provides an analysis of the data structure and gives meaning to the input data for data generation. Experimental results show that approximating data distributions using hierarchical structures not only facilitates the observation of data distributions and the control of data generation but also greatly improves data generation and data distribution approximation performance.