Abstract: Traditional Neural Networks often ignore the underlying geometry of data, treating inputs like images or graphs as simple lists of numbers. In this presentation, we explore how Geometric Deep Learning (GDL) solves this by building mathematical symmetries directly into the network’s architecture. This approach allows models to learn much more efficiently, as they don’t need to see every possible orientation of an object to recognize it. Starting from the sliding window of Convolutional Neural Networks (CNNs), we will explain how to design filters that remain effective even when data is rotated or scaled. We will then introduce techniques such as steerable filters and equivariant transformers, demonstrating their practical use in 3D point clouds.