Applied math seminar - Dr. Shay Deutsch

Speaker
Dr. Shay Deutsch
Date
26/12/2019 - 15:00 - 14:00Add to Calendar 2019-12-26 14:00:00 2019-12-26 15:00:00 Applied math seminar - Dr. Shay Deutsch Title: Robust Methods for Topology Estimation in Unsupervised Learning Abstract: Learning graph connectivity has broad-ranging applications from 3D reconstruction to unsupervised learning. In this talk I will introduce a new method to learn the graph structure underlying noisy point set observations assumed to lie near a complex manifold. Rather than assuming regularity of the manifold itself, as customary, we assume regularity of the geodesic flow through the boundary of arbitrary regions on the graph. The idea is to exploit this more flexible notion of regularity, captured by the discrete equivalent of the isoperimetric inequality for closed manifolds, to infer the graph structure.. In a broader perspective, when studying the topology of the graph networks, we would like to learn new representations that capture not only local connectivity, i.e., nodes that belong to the same local structure, but also similarity which is based on their structural role in the graph. I will discuss a new approach and vision towards learning a good trade-off between these local and structural types of similarities that includes diverse possible applications including point clouds, biological networks and social networks. Bldg. 216, Room 201 אוניברסיטת בר-אילן - המחלקה למתמטיקה mathoffice@math.biu.ac.il Asia/Jerusalem public
Place
Bldg. 216, Room 201
Abstract

Title: Robust Methods for Topology Estimation in Unsupervised Learning

Abstract:

Learning graph connectivity has broad-ranging applications from 3D reconstruction to unsupervised learning. In this talk I will introduce a new method to learn the graph structure underlying noisy point set observations assumed to lie near a complex manifold. Rather than assuming regularity of the manifold itself, as customary, we assume regularity of the geodesic flow through the boundary of arbitrary regions on the graph. The idea is to exploit this more flexible notion of regularity, captured by the discrete equivalent of the isoperimetric inequality for closed manifolds, to infer the graph structure..

In a broader perspective, when studying the topology of the graph networks, we would like to learn new representations that capture not
only local connectivity, i.e., nodes that belong to the same local structure, but also similarity which is based on their structural role in the graph. I will discuss a new approach and vision towards learning a good trade-off between these local and structural types of similarities that includes diverse possible applications including point clouds, biological networks and social networks.

תאריך עדכון אחרון : 22/12/2019