Prof. Dr. Florian Lemmerich, leader of the research group of "Applied Machine Learning" at the University of Passau, will give a presentation on the topic
"Same data, same algorithm, same results? On the predictive stability of training Deep Learning models"
as part of the VAC Colloquium.
Afterwards we look forward to discussions while enjoying coffee and cookies.
The event is open to anyone interested.
Abstract:
Deep Learning is nowadays the go-to solution for many machine learning tasks. However, the inherent randomness within the training process introduces a curious phenomenon: basically equivalent or even the exact same training data with the same algorithm and parameterization may yield disparate outcomes. Despite achieving comparable aggregate performance, individual predictions and internal representations can exhibit significant variance. In this talk, we give a short introduction into this phenomenon from an empirical perspective, covering potential sources of instability, challenges in measuring, and potential application consequences.