In this session we will discuss the relationship between data augmentation, invariant/equivariant features and the abstract concept of convolution layers, specifically, how this abstraction can be extended to devise concrete neural network architectures that are robust to diverse data and augmentation types (all of which are published after 2016). We will focus on the latest of the series, designed to handle spatial graph data augmentable by 3d translations & rotations. In the end, we will showcase it’s applications in molecule analysis and autonomous flight. We will focus on the latest of the series, designed to handle spatial graph data augmentable by 3d translations & rotations. In the end, we will showcase it’s applications in molecule analysis and autonomous flight.