Title: Efficient Continuous Symmetry Discovery and Enforcement in Geometric AI
Abstract: Symmetry-aware machine learning has recently gained interest in data science and AI/ML. Discovering symmetries of a machine learning model can provide insight about anticipated behavior on previously unseen data. Meanwhile, enforcing model symmetries can result in more favorable model performance, whether the symmetries are learned from data or specified in advance. However, current methods for continuous symmetry discovery and enforcement are computationally expensive, and the types of symmetries that can be considered are typically restricted. This talk will show how the Lie derivative can be leveraged to engage with more complicated symmetries and in a computationally efficient manner. Applications to computer vision will be highlighted.