Thema der Dissertation:
Geometric Data Analysis: Advancements of the Statistical Methodology and Applications Thema der Disputation:
Convolutions over non-Euclidean Domains
Geometric Data Analysis: Advancements of the Statistical Methodology and Applications Thema der Disputation:
Convolutions over non-Euclidean Domains
Abstract: Euclidean convolutions have been studied for a long time, and they have found numerous applications, e.g., in physics and engineering. More recently, they have been used in machine learning tasks and are now an integral part of many deep-learning architectures. Motivated by complex learning problems, several ways to generalize convolutions for non-Euclidean domains like (discrete) manifolds [1, 3] and graphs [1, 2] have been investigated. When used as part of the architecture, they can help to increase the performance. In this talk, we discuss some of these generalizations–how they are motivated and which properties of the Euclidean convolution they share–and show selected applications.
[1] https://arxiv.org/pdf/2104.13478.pdf (mainly Chapter 4)
[2] https://arxiv.org/pdf/1606.09375.pdf
[3] https://par.nsf.gov/servlets/purl/10336824.
[2] https://arxiv.org/pdf/1606.09375.pdf
[3] https://par.nsf.gov/servlets/purl/10336824.
Zeit & Ort
13.07.2023 | 11:30
Seminarraum 2006
(Zuse Institut Berlin, Takustr.7,14195 Berlin)