On the Geometry of Rectifier Convolutional Neural Networks

Matteo Gamba, Hossein Azizpour, Stefan Carlsson, Mårten Björkman

ICCVW 2019 pp. 793-797

doi:10.1109/ICCVW.2019.00106 /iccvw/2019/gamba2019iccvw-geometry/

Abstract

While recent studies have shed light on the expressivity, complexity and compositionality of convolutional networks, the real inductive bias of the family of functions reachable by gradient descent on natural data is still unknown. By exploiting symmetries in the preactivation space of convolutional layers, we present preliminary empirical evidence of regularities in the preimage of trained rectifier networks, in terms of arrangements of polytopes, and relate it to the nonlinear transformations applied by the network to its input.

ICCVW Semantic Scholar

Cite

Text

Gamba et al. "On the Geometry of Rectifier Convolutional Neural Networks." IEEE/CVF International Conference on Computer Vision Workshops, 2019. doi:10.1109/ICCVW.2019.00106

Markdown

[Gamba et al. "On the Geometry of Rectifier Convolutional Neural Networks." IEEE/CVF International Conference on Computer Vision Workshops, 2019.](https://mlanthology.org/iccvw/2019/gamba2019iccvw-geometry/) doi:10.1109/ICCVW.2019.00106

BibTeX

@inproceedings{gamba2019iccvw-geometry,
  title     = {{On the Geometry of Rectifier Convolutional Neural Networks}},
  author    = {Gamba, Matteo and Azizpour, Hossein and Carlsson, Stefan and Björkman, Mårten},
  booktitle = {IEEE/CVF International Conference on Computer Vision Workshops},
  year      = {2019},
  pages     = {793-797},
  doi       = {10.1109/ICCVW.2019.00106},
  url       = {https://mlanthology.org/iccvw/2019/gamba2019iccvw-geometry/}
}