Get Started
Home
Topics
Search
Library
Research questionHow can VAEs reparameterize latent distributions on manifolds with nontrivial topology?Standard VAE reparameterizations are designed for simple latent spaces, making optimization and KL-divergence computation difficult when the latent manifold has nontrivial topology.
AI
Computer Vision
Machine Learning
Statistical Machine Learning
Latest papersRecent research connected to this question, newest first.Reparameterization through Coverings and Topological Weight PriorsThe work develops reparameterization through covering maps and derives a KL-divergence bound relating distributions on a base manifold to those on its cover. It demonstrates the approach with a VAE whose latent space has Klein-bottle topology and discusses topology-informed weight priors for Bayesian convolutional vision models.research paper · Sep 1, 2026
Related questions
How can PAC-Bayesian reconstruction guarantees for variational autoencoders handle temporal dependence in time-series data?Can stationary-point ELBO values be expressed as entropy sums across variational generative models?How can explicit surface representations handle varied topologies while preserving global smoothness and differentiable geometric queries?How can SPD-manifold representations capture intrinsic relational geometry while preserving valid matrix constraints?