Timezone: »
To better conform to data geometry, recent deep generative modelling techniques adapt Euclidean constructions to non-Euclidean spaces. In this paper, we study normalizing flows on manifolds. Previous work has developed flow models for specific cases; however, these advancements hand craft layers on a manifold-by-manifold basis, restricting generality and inducing cumbersome design constraints. We overcome these issues by introducing Neural Manifold Ordinary Differential Equations, a manifold generalization of Neural ODEs, which enables the construction of Manifold Continuous Normalizing Flows (MCNFs). MCNFs require only local geometry (therefore generalizing to arbitrary manifolds) and compute probabilities with continuous change of variables (allowing for a simple and expressive flow construction). We find that leveraging continuous manifold dynamics produces a marked improvement for both density estimation and downstream tasks.
Author Information
Aaron Lou (Cornell University)
Derek Lim (Cornell University)
Isay Katsman (Cornell University)
Leo Huang (Cornell University)
Qingxuan Jiang (Cornell University)
Ser Nam Lim (Facebook AI)
Christopher De Sa (Cornell)
More from the Same Authors
-
2020 Workshop: Differential Geometry meets Deep Learning (DiffGeo4DL) »
Joey Bose · Emile Mathieu · Charline Le Lan · Ines Chami · Frederic Sala · Christopher De Sa · Maximillian Nickel · Christopher RĂ© · Will Hamilton -
2020 Poster: Better Set Representations For Relational Reasoning »
Qian Huang · Horace He · Abhay Singh · Yan Zhang · Ser Nam Lim · Austin Benson -
2020 Poster: Random Reshuffling is Not Always Better »
Christopher De Sa -
2020 Poster: Asymptotically Optimal Exact Minibatch Metropolis-Hastings »
Ruqi Zhang · A. Feder Cooper · Christopher De Sa -
2020 Spotlight: Asymptotically Optimal Exact Minibatch Metropolis-Hastings »
Ruqi Zhang · A. Feder Cooper · Christopher De Sa -
2020 Spotlight: Random Reshuffling is Not Always Better »
Christopher De Sa -
2019 Poster: Numerically Accurate Hyperbolic Embeddings Using Tiling-Based Models »
Tao Yu · Christopher De Sa -
2019 Spotlight: Numerically Accurate Hyperbolic Embeddings Using Tiling-Based Models »
Tao Yu · Christopher De Sa -
2019 Poster: Dimension-Free Bounds for Low-Precision Training »
Zheng Li · Christopher De Sa -
2019 Poster: Poisson-Minibatching for Gibbs Sampling with Convergence Rate Guarantees »
Ruqi Zhang · Christopher De Sa -
2019 Spotlight: Poisson-Minibatching for Gibbs Sampling with Convergence Rate Guarantees »
Ruqi Zhang · Christopher De Sa -
2019 Poster: Channel Gating Neural Networks »
Weizhe Hua · Yuan Zhou · Christopher De Sa · Zhiru Zhang · G. Edward Suh