The shape of adversarial training by Nicolás García Trillos
Date: April 19, 2023
Speaker: Nicolás García Trillos
Title: The shape of adversarial training
Abstract: This talk is about two apparently non-overlapping stories. One story is about shapes in space (Euclidean space or a network), their perimeter, and their curvature. The other story is about machine learning, specifically about how to train learning models to be robust to adversarial perturbations of data. The bigger story in the talk will be about how these two stories interact with each other, how adversarial robustness motivates new notions of perimeter and curvature, and how geometry can cast new lights on and in this way reveal new faces of an important task in machine learning.
The talk is based on several recent works, some in progress, with multiple collaborators, including my PhD student Jakwang Kim, with whom I have worked on several aspects of this line of research. The discussion will be informal and will be mostly focused on conveying high-level ideas and concepts. No background in geometry is required.