Group Equivariant Convolutional Networks

Maya Janvier

Keywords: Machine learning, group geometry, GAN, Higgs Boson

TL;DR: Discover Group Equivariant CNN properties, performances on datasets, GAN extension and its applications in Physics

Abstract:

Convolutional Neural Networks (CNN) are using symmetry priors to make the
best out of the properties of the data, in particular translation invariance in images.
Group Equivariant CNN (Cohen & Welling, 2016) extend CNN by using invari-
ances from other groups of symmetry. After exploring their mathematical proper-
ties, we confirm their performances on Rotated MNIST and CIFAR10, introduce
some extensions with GAN and propose applications in High Energy Physics.

Community Implementations: [ 2 code implementations](https://www.catalyzex.com/paper/arxiv:1602.07576/code)