Our colloquium series offers a diverse platform for research scholars, faculty, students, and industry experts to share and exchange ideas, fostering discussion and networking across mathematics, statistics and data science.
Abiy Tasissa of Tufts University will speak at this week's colloquium on "Learning Geometry From Anchored Data."
Abstract: Many learning problems require inferring global structure from incomplete, local or relational data. We study this problem through the lens of anchors: selected points that provide partial geometric or relational information about the rest of the data, and serve as an inductive bias for learning. We develop this perspective in two settings. First, we consider representing data as convex combinations of local exemplars, where both the exemplars (anchors) and the representations are learned from data. This leads to sparse, structured representations and reveals connections to structured compressive sensing and to neural architectures with built-in geometric structure. Second, we study the problem of recovering point locations from partial distance measurements to anchors. Unlike standard localization and graph-based methods, we do not assume distances between non-anchor points, nor complete distance information among anchors. We present an optimization-based approach for estimating global geometry under these minimal assumptions. We conclude by noting that anchors can serve as a unifying framework for learning geometry and structure from partial information, and that they highlight applications in resource-constrained sensor networks, structure prediction, manifold learning and interpretable deep learning.
Speaker Bio: Abiy Tasissa received a B.Sc. in mathematics and an M.Sc. in aeronautics and astronautics from the Massachusetts Institute of Technology. She also earned a doctorate in applied mathematics from Rensselaer Polytechnic Institute. He is currently an assistant professor in the Department of Mathematics at Tufts University. His research focuses on developing provable algorithms to estimate structures from incomplete distance data. He is also interested in provable algorithms for signal processing and statistical learning, structured deep learning and applied linear algebra.
Read More