“…First, the organization of visual data based on their location and orientation [Hubel and Wiesel, 1962] is modeled by Lie group convolutions [Bekkers, 2019], in which feature maps encode response for every position and every orientation. Second, long-range interactions between aligned neurons [Bosking et al, 1997] are modeled by building graphs with affinity matrices based on (approximate) sub-Riemannian distances on the Lie groups, inspired by sub-Riemannian image analysis methods such as [Franken and Duits, 2009, Favali et al, 2016, Mashtakov et al, 2017, Boscain et al, 2018, Duits et al, 2018, Baspinar et al, 2021. Defferrard et al [2020] showed how to construct powerful graph NNs that are faithful to the manifolds on which they are defined.…”