Meet DTU Compute: Professor Aasa Feragen

“Meet DTU Compute” is a lecture series giving everybody the opportunity to hear good colleagues from DTU Compute give friendly introductions to hot research topics. The next Meet DTU Compute lecture will be given by professor Aasa Feragen Tuesday, May 19 at 14:00 - 15:00.

 

Speaker: Professor Aasa Feragen, Section for Image Analysis & Computer Graphics, DTU Compute

 

Title: Uncertainty quantification in non-Euclidean domains

Abstract: While raw data is usually handed to us in the form of Euclidean vectors of measurements, incorporation of prior or learned knowledge often leads us to leave the Euclidean domain: This can happen via enforcing known constraints or invariances that the data should adhere to, or by learning spatially varying dependencies and structures in the data, for instance via metric learning.

 

In the first part of this talk, we will explain how Riemannian manifolds naturally appear as a data domain when incorporating prior or learned knowledge, and we will discuss basic properties of Riemannian manifolds that allow us to define Gaussian distributions and Gaussian processes on them. In the second part of the talk, we will showcase how manifold-valued Gaussian processes can be used to quantify uncertainty in manifold-valued regression and submanifold learning. Finally, we will discuss how these models relate to current geometry-aware models and open problems in deep learning.

 

You can join Meet DTU Compute lecture online at zoom here: https://dtudk.zoom.us/j/63473605968

 

Everyone is welcome!

Time

Tue 19 May 20
14:00 - 15:00

Organizer

DTU Compute

Where

The lecture will take place in ZOOM, here:

https://dtudk.zoom.us/j/63473605968