Explore a video lecture on gauge equivariant convolutional networks and their application to icosahedral CNNs. Delve into the extension of equivariance principles from global symmetries to local gauge transformations, enabling the development of a general class of convolutional neural networks on manifolds. Learn how this approach depends only on intrinsic geometry and encompasses various methods from equivariant and geometric deep learning. Discover the implementation of gauge equivariant CNNs for signals on the icosahedron surface, offering a practical approximation of spherical data. Understand how this method achieves significant improvements in tasks such as segmenting omnidirectional images and global climate patterns, demonstrating its potential as a scalable alternative to Spherical CNNs.
Overview
Syllabus
Gauge Equivariant Convolutional Networks and the Icosahedral CNN
Taught by
Yannic Kilcher