Abstract
PDE-based Group Convolutional Neural Networks (PDE-G-CNNs) use solvers ofevolution PDEs as substitutes for the conventional components in G-CNNs.PDE-G-CNNs can offer several benefits simultaneously: fewer parameters,inherent equivariance, better accuracy, and data efficiency. In this article we focus on Euclidean equivariant PDE-G-CNNs where thefeature maps are two-dimensional throughout. We call this variant of theframework a PDE-CNN. From a machine learning perspective, we list several practically desirableaxioms and derive from these which PDEs should be used in a PDE-CNN, this beingour main contribution. Our approach to geometric learning via PDEs is inspiredby the axioms of scale-space theory, which we generalize by introducingsemifield-valued signals. Our theory reveals new PDEs that can be used in PDE-CNNs and weexperimentally examine what impact these have on the accuracy of PDE-CNNs. Wealso confirm for small networks that PDE-CNNs offer fewer parameters, increasedaccuracy, and better data efficiency when compared to CNNs.