Catherine Potts (Dept. of Mathematical Sciences, MSU) 

11/16/2021  12:00pm

Abstract:

In this talk, we will take an informal look at adapting archetypal analysis to large-scale imaging problems that crop up in scientific investigations. The geometric interpretation of archetypal analysis will be the basis for the discussion, and how the geometry can be leveraged to manage these large-scale high-dimensional data collections. Being an informal discussion, the algebraic formulations will not be presented, instead we will take an intuitive journey through the applications and the components of this unsupervised machine learning method. I will also present the adaptation developed as part of my dissertation, which is a sketch-based algorithm that lowers the data dimension while maintaining the geometric aspects of the data cloud that are vital for archetypal analysis.


Mathematical concepts discussed will include constrained optimization, machine learning, imaging problems, and convex geometry. Applications presented will include neuroscience imaging and exoplanet classifications, as well as some classical computer vision imaging datasets.