Recent work (Cohen & Welling, 2016) has shown that generalizations of
convolutions, based on group theory, provide powerful inductive biases for
learning. In these generalizations, filters are not only translated but can
also be rotated, flipped, etc. However, coming up with exact mode