Abstract: When using neural networks to approximate functions in lattice gauge theory, the networks should (1) transform correctly under gauge transformations and (2) preserve any (global) lattice symmetries. These properties increase robustness and reduce the need for data augmentation. Whereas group-equivariant convolutional neural networks are known to preserve global symmetry, they do not preserve gauge symmetry. In this seminar we discuss how convolutional layers can be modified to also make them gauge equivariant.