More Like this

1. Principal Component Flows

   Cunningham E.; Cobb A.; Jha, S (April 2022, Proceedings of the 39th International Conference on Machine Learning, PMLR 162:4492-4519)

   Normalizing flows map an independent set of latent variables to their samples using a bijective transformation. Despite the exact correspondence between samples and latent variables, their high level relationship is not well understood. In this paper we characterize the geometric structure of flows using principal manifolds and understand the relationship between latent variables and samples using contours. We introduce a novel class of normalizing flows, called principal component flows (PCF), whose contours are its principal manifolds, and a variant for injective flows (iPCF) that is more efficient to train than regular injective flows. PCFs can be constructed using any flow architecture, are trained with a regularized maximum likelihood objective and can perform density estimation on all of their principal manifolds. In our experiments we show that PCFs and iPCFs are able to learn the principal manifolds over a variety of datasets. Additionally, we show that PCFs can perform density estimation on data that lie on a manifold with variable dimensionality, which is not possible with existing normalizing flows. 

   more » « less
2. VQ-Flows: Vector Quantized Local Normalizing Flows

   Sidheekh, Sahil; Dock, Chris B.; Jain, Tushar; Balan, Radu; Singh, Maneesh K. (August 2022, Uncertainty in artificial intelligence)

   Normalizing flows provide an elegant approach to generative modeling that allows for efficient sampling and exact density evaluation of unknown data distributions. However, current techniques have significant limitations in their expressivity when the data distribution is supported on a lowdimensional manifold or has a non-trivial topology. We introduce a novel statistical framework for learning a mixture of local normalizing flows as “chart maps” over the data manifold. Our framework augments the expressivity of recent approaches while preserving the signature property of normalizing flows, that they admit exact density evaluation. We learn a suitable atlas of charts for the data manifold via a vector quantized autoencoder (VQ-AE) and the distributions over them using a conditional flow. We validate experimentally that our probabilistic framework enables existing approaches to better model data distributions over complex manifolds. 

   more » « less
3. Maximally highly proximal flows

   https://doi.org/10.1017/etds.2020.40  

   ZUCKER, ANDY (May 2020, Ergodic Theory and Dynamical Systems)

   For G a Polish group, we consider G-flows which either contain a comeager orbit or have all orbits meager. We single out a class of flows, the maximally highly proximal (MHP) flows, for which this analysis is particularly nice. In the former case, we provide a complete structure theorem for flows containing comeager orbits, generalizing theorems of Melleray, Nguyen Van Thé, and Tsankov and of Ben Yaacov, Melleray, and Tsankov. In the latter, we show that any minimal MHP flow with all orbits meager has a metrizable factor with all orbits meager, thus ‘reflecting’ complicated dynamical behavior to metrizable flows. We then apply this to obtain a structure theorem for Polish groups whose universal minimal flow is distal. 

   more » « less
4. Normalizing Flows on Tori and Spheres

   Rezende, Danilo Jimenez; Papamakarios, George; Racaniere, Sebastian; Albergo, Michael; Kanwar, Gurtej; Shanahan, Phiala; Cranmer, Kyle (January 2020, Proceedings of Machine Learning Research)

   null (Ed.)

   Normalizing flows are a powerful tool for building expressive distributions in high dimensions. So far, most of the literature has concentrated on learning flows on Euclidean spaces. Some problems however, such as those involving angles, are defined on spaces with more complex geometries, such as tori or spheres. In this paper, we propose and compare expressive and numerically stable flows on such spaces. Our flows are built recursively on the dimension of the space, starting from flows on circles, closed intervals or spheres. 

   more » « less
5. Hopper flows of deformable particles

   https://doi.org/10.1039/d2sm01079h  

   Cheng, Yuxuan; Treado, John D.; Lonial, Benjamin F.; Habdas, Piotr; Weeks, Eric R.; Shattuck, Mark D.; O'Hern, Corey S. (November 2022, Soft Matter)

   Numerous experimental and computational studies show that continuous hopper flows of granular materials obey the Beverloo equation that relates the volume flow rate Q and the orifice width w : Q ∼ ( w / σ avg − k ) β , where σ avg is the average particle diameter, kσ avg is an offset where Q ∼ 0, the power-law scaling exponent β = d − 1/2, and d is the spatial dimension. Recent studies of hopper flows of deformable particles in different background fluids suggest that the particle stiffness and dissipation mechanism can also strongly affect the power-law scaling exponent β . We carry out computational studies of hopper flows of deformable particles with both kinetic friction and background fluid dissipation in two and three dimensions. We show that the exponent β varies continuously with the ratio of the viscous drag to the kinetic friction coefficient, λ = ζ / μ . β = d − 1/2 in the λ → 0 limit and d − 3/2 in the λ → ∞ limit, with a midpoint λ c that depends on the hopper opening angle θ w . We also characterize the spatial structure of the flows and associate changes in spatial structure of the hopper flows to changes in the exponent β . The offset k increases with particle stiffness until k ∼ k max in the hard-particle limit, where k max ∼ 3.5 is larger for λ → ∞ compared to that for λ → 0. Finally, we show that the simulations of hopper flows of deformable particles in the λ → ∞ limit recapitulate the experimental results for quasi-2D hopper flows of oil droplets in water. 

   more » « less