Flows for simultaneous manifold learning and density estimation
J. Brehmer, and K. Cranmer. (2020)cite arxiv:2003.13913Comment: Code at https://github.com/johannbrehmer/manifold-flow.
Abstract
We introduce manifold-modeling flows (MFMFs), a new class of generative
models that simultaneously learn the data manifold as well as a tractable
probability density on that manifold. Combining aspects of normalizing flows,
GANs, autoencoders, and energy-based models, they have the potential to
represent data sets with a manifold structure more faithfully and provide
handles on dimensionality reduction, denoising, and out-of-distribution
detection. We argue why such models should not be trained by maximum likelihood
alone and present a new training algorithm that separates manifold and density
updates. With two pedagogical examples we demonstrate how manifold-modeling
flows let us learn the data manifold and allow for better inference than
standard flows in the ambient data space.
Description
[2003.13913] Flows for simultaneous manifold learning and density estimation
%0 Journal Article
%1 brehmer2020flows
%A Brehmer, Johann
%A Cranmer, Kyle
%D 2020
%K flows generative-models
%T Flows for simultaneous manifold learning and density estimation
%U http://arxiv.org/abs/2003.13913
%X We introduce manifold-modeling flows (MFMFs), a new class of generative
models that simultaneously learn the data manifold as well as a tractable
probability density on that manifold. Combining aspects of normalizing flows,
GANs, autoencoders, and energy-based models, they have the potential to
represent data sets with a manifold structure more faithfully and provide
handles on dimensionality reduction, denoising, and out-of-distribution
detection. We argue why such models should not be trained by maximum likelihood
alone and present a new training algorithm that separates manifold and density
updates. With two pedagogical examples we demonstrate how manifold-modeling
flows let us learn the data manifold and allow for better inference than
standard flows in the ambient data space.
@article{brehmer2020flows,
abstract = {We introduce manifold-modeling flows (MFMFs), a new class of generative
models that simultaneously learn the data manifold as well as a tractable
probability density on that manifold. Combining aspects of normalizing flows,
GANs, autoencoders, and energy-based models, they have the potential to
represent data sets with a manifold structure more faithfully and provide
handles on dimensionality reduction, denoising, and out-of-distribution
detection. We argue why such models should not be trained by maximum likelihood
alone and present a new training algorithm that separates manifold and density
updates. With two pedagogical examples we demonstrate how manifold-modeling
flows let us learn the data manifold and allow for better inference than
standard flows in the ambient data space.},
added-at = {2020-04-01T18:53:10.000+0200},
author = {Brehmer, Johann and Cranmer, Kyle},
biburl = {https://www.bibsonomy.org/bibtex/268902625444f7a769261b02975fc980a/kirk86},
description = {[2003.13913] Flows for simultaneous manifold learning and density estimation},
interhash = {e90c140579995af1b76c083315f1a83e},
intrahash = {68902625444f7a769261b02975fc980a},
keywords = {flows generative-models},
note = {cite arxiv:2003.13913Comment: Code at https://github.com/johannbrehmer/manifold-flow},
timestamp = {2020-04-01T18:53:10.000+0200},
title = {Flows for simultaneous manifold learning and density estimation},
url = {http://arxiv.org/abs/2003.13913},
year = 2020
}