Seminar: Comparing CNNs with Group Convolution-based Rotationally Invariant CNNs on Storm Data

Speaker

Lander Ver Hoef, Colorado State University

Bio

Lander Ver Hoef (he/him/his) is a mathematics PhD student from Colorado State University. His research interests include geometric deep learning, topological data analysis, and applied harmonic analysis. Prior to graduate school, he worked as an officer on a NOAA hydrographic research ship doing both ship handling and operations as well as collecting and processing hydrographic data.

Abstract

Group convolution provides a framework for creating neural networks that respect certain symmetries of the input space. We began our project under the hypothesis that a network that was invariant to rotations would be more efficient and generalizable on storm data. In this talk we’ll provide an introduction to geometric deep learning and group convolution, as well as discuss in what ways this hypothesis holds true and in what ways it doesn’t. We’ll also explore differences in how these approaches respond to explainable AI techniques.

Add this event to your calendar (internal only).