Abstract: Convolutional neural networks (CNNs) have achieved remarkable empirical success on a wide range of highly complex tasks. A possible explanation for this success is the way traditional CNNs respect translational symmetry in the input data, e.g. images, to extract efficient and meaningful representations of it. In geometric deep learning, symmetries of data are the fundamental principle underlying the construction of models generalizing CNNs to data defined on more general manifolds and exhibiting more general groups of symmetries. We discuss one class of such models, known as group equivariant convolutional neural networks (GCNNs), based on the recent review paper arXiv:2105.13926 [cs.LG].
To receive the Zoom link, please contact the seminar organiser: Niklas Lundström
