In deep learning and computer vision, it is common for data to present certain symmetries. For instance, histopathological scans, satellite images or protein structures can have an arbitrary orientation or, in natural scenes, objects can freely rotate around their vertical axis. This prior knowledge about the symmetries of a problem can be leveraged via an equivariant model design. In particular, steerable CNNs are one of the most common and general classes of equivariant neural networks. In this talk, we will review the framework of steerable CNNs and present a theoretical characterization of general steerable kernel spaces and a practical program to parameterize steerable filters. Our theory enables us to directly parameterize filters in terms of a band-limited basis on the Euclidean space and to easily implement steerable CNNs equivariant to a large number of groups. These include new architectures equivariant to, for example, the symmetries of the platonic solids or to 3D azimuthal symmetries (rotations around the Z axis).
Speaker: Gabriele Cesa (Website)
About the speaker: Gabriele Cesa is a Research Associate at Qualcomm AI Research, Amsterdam and a Ph.D. student at AMLab, University of Amsterdam, under the supervision of Max Welling. Gabriele’s research focuses on augmenting machine learning methods with prior information about the geometry of a problem to achieve improved data efficiency and generalization. A particular emphasis has been given to equivariant neural networks, which can encode our knowledge about the symmetries in the data into the model’s architecture. In the last years, Gabriele developed the Python library “e2cnn” (now “escnn”), used by many practitioners to implement steerable CNNs.
Previously, Gabriele received a Master’s degree in Artificial Intelligence at the University of Amsterdam and a Bachelor’s degree in Computer Science at the University of Trento.