Title: Do Deep Ensembles Capture Uncertainty in Graph Neural Networks?
 
Abstract: Deep ensembles are the standard for uncertainty quantification in deep learning, but their effectiveness for graph-structured data is often just assumed based on their success in other domains, like computer vision. This talk presents the results of my recent paper, which studies deep ensembles specifically for message-passing graph neural networks. We show that ensembles provide surprisingly little improvement over a single model for uncertainty quantification, and we investigate this phenomenon.

Speaker: Viacheslav Borovitskiy is a Lecturer in Machine Learning at the University of Edinburgh. His research lies at the intersection of mathematics and machine learning, with a focus on geometric learning and uncertainty quantification. His contributions have been recognized with best-paper-type awards at top-tier machine learning conferences. Previously, he was an ETH Fellow at ETH Zürich and received his PhD in Harmonic Analysis from the Steklov Mathematical Institute.

Livestream: https://www.chaspark.com/#/live/1283555091265658880?multi=en

【Tech Talk】Do Deep Ensembles Capture Uncertainty in Graph Neural Networks? / Huawei-Edinburgh Joint Lab by blogadmin is licensed under a Creative Commons Attribution CC BY 3.0