Explorations in Homeomorphic Variational Auto-Encoding

Abstract

The manifold hypothesis states that many kinds of high-dimensional data are concentrated near a low-dimensional manifold. If the topology of this data manifold is non-trivial, a continuous encoder network cannot embed it in a one-to-one manner without creating holes of low density in the latent space. This is at odds with the Gaussian prior assumption typically made in Variational Auto-Encoders (VAEs), because the density of a Gaussian concentrates near a blob-like manifold. In this paper we investigate the use of manifold-valued latent variables. Specifically, we focus on the important case of continuously differentiable symmetry groups (Lie groups), such as the group of 3D rotations $\mathrm{SO}(3)$. We show how a VAE with $\mathrm{SO}(3)$-valued latent variables can be constructed, by extending the reparameterization trick to compact connected Lie groups. Our experiments show that choosing manifold-valued latent variables that match the topology of the latent data manifold, is crucial to preserve the topological structure and learn a well-behaved latent space.

arXiv comment16 pages, 8 figures, ICML workshop on Theoretical Foundations and Applications of Deep Generative Models

References in corpus (2)

• Intertwiners between Induced Representations (with Applications to the Theory of Equivariant Neural Networks)
• Harmonic Exponential Families on Manifolds

Cited by in corpus (34)

• What is a meaningful representation of protein sequences?
• Data Augmentation in High Dimensional Low Sample Size Setting Using a Geometry-Based Variational Autoencoder
• Deep generative models in inversion: a review and development of a new approach based on a variational autoencoder
• Normalizing Flows on Tori and Spheres
• On the Continuity of Rotation Representations in Neural Networks
• Generative Capacity of Probabilistic Protein Sequence Models
• Learning the Stein Discrepancy for Training and Evaluating Energy-Based Models without Sampling
• Diffusion Variational Autoencoders
• Stochastic Normalizing Flows
• The Power Spherical distribution
• Neural Ordinary Differential Equations on Manifolds
• Variational Autoencoders for Learning Nonlinear Dynamics of Physical Systems
• Coordinate Independent Convolutional Networks -- Isometry and Gauge Equivariant Convolutions on Riemannian Manifolds
• Variational Autoencoders with Riemannian Brownian Motion Priors
• Chart Auto-Encoders for Manifold Structured Data
• Symmetry and Complexity in Object-Centric Deep Active Inference Models
• Variational Integrator Networks for Physically Structured Embeddings
• Pythae: Unifying Generative Autoencoders in Python -- A Benchmarking Use Case
• Semantics Preserving Adversarial Learning
• Manifolds for Unsupervised Visual Anomaly Detection
• A Neural Rendering Framework for Free-Viewpoint Relighting
• Spherical Regression: Learning Viewpoints, Surface Normals and 3D Rotations on n-Spheres
• Increasing Expressivity of a Hyperspherical VAE
• On the Latent Holes of VAEs for Text Generation
• Disentanglement with Hyperspherical Latent Spaces using Diffusion Variational Autoencoders
• No Representation without Transformation
• On the Generative Utility of Cyclic Conditionals
• End-to-End Simultaneous Learning of Single-particle Orientation and 3D Map Reconstruction from Cryo-electron Microscopy Data
• Learning Equivariant Representations
• Joint Estimation of Image Representations and their Lie Invariants
• Embedding-reparameterization procedure for manifold-valued latent variables in generative models
• Hyperbolic Manifold Regression
• Deep Autoencoders: From Understanding to Generalization Guarantees
• Autoequivariant Network Search via Group Decomposition