steerable machine learning

Inspired by the hope of mashing up our paper on scalars and vectors in machine learning with work like this on steerable neural networks, I worked with Soledad Villar (JHU) today on writing down all possible linear convolutional kernels in a 3x3x3 block of a 3-d image that satisfy the geometric symmetries of scalar, vector, and second-order tensor forms. I feel like recent work on machine-learning in cosmology like this could be vastly improved by these geometric methods.