Embedding-reparameterization procedure for manifold-valued latent variables in generative models

2018/12/06 by Eugène Golikov, Eugene Golikov, Golikov, Eugene +2
Computer Science · Mathematics · #FOS: Computer and information sciences #Gaussian Processes and Bayesian Inference #Generative Adversarial Networks and Image Synthesis #Machine Learning (cs.LG) #Machine Learning (stat.ML) #Music and Audio Processing #cs.LG #stat.ML

paper · pdf · doi:10.48550/arxiv.1812.02769

Presented at Bayesian Deep Learning workshop (NeurIPS 2018)

arxiv created 2018/12/06 · openalex publication_date 2018/12/06 · arxiv updated 2018/12/10 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28

Abstract

Conventional prior for Variational Auto-Encoder (VAE) is a Gaussian distribution. Recent works demonstrated that choice of prior distribution affects learning capacity of VAE models. We propose a general technique (embedding-reparameterization procedure, or ER) for introducing arbitrary manifold-valued variables in VAE model. We compare our technique with a conventional VAE on a toy benchmark problem. This is work in progress.

Citations

• Importance Weighted Autoencoders
• Generative Adversarial Networks: An Overview
• Spherical Latent Spaces for Stable Variational Autoencoders

Related

• Generative Model without Prior Distribution Matching
• Variational Autoencoder with Learned Latent Structure
• A Geometric Perspective on Variational Autoencoders
• Hyperbolic VAE via Latent Gaussian Distributions
• From Points to Spheres: A Geometric Reinterpretation of Variational Autoencoders