English

Generalization of the Change of Variables Formula with Applications to Residual Flows

Machine Learning 2021-07-12 v1 Machine Learning Dynamical Systems

Abstract

Normalizing flows leverage the Change of Variables Formula (CVF) to define flexible density models. Yet, the requirement of smooth transformations (diffeomorphisms) in the CVF poses a significant challenge in the construction of these models. To enlarge the design space of flows, we introduce L\mathcal{L}-diffeomorphisms as generalized transformations which may violate these requirements on zero Lebesgue-measure sets. This relaxation allows e.g. the use of non-smooth activation functions such as ReLU. Finally, we apply the obtained results to planar, radial, and contractive residual flows.

Keywords

Cite

@article{arxiv.2107.04346,
  title  = {Generalization of the Change of Variables Formula with Applications to Residual Flows},
  author = {Niklas Koenen and Marvin N. Wright and Peter Maaß and Jens Behrmann},
  journal= {arXiv preprint arXiv:2107.04346},
  year   = {2021}
}