14 May 2025
Rosensäle
Europe/Berlin timezone

Towards Reliable AI: Applications of Uncertainty Quantification in Deep Learning

Not scheduled
20m
Rosensäle

Rosensäle

Fürstengraben 27 · 07743 Jena

Description

As deep learning systems are increasingly deployed in high-stakes domains such as medical diagnostics, climate modeling, and autonomous decision-making, their ability to express uncertainty in their predictions becomes crucial. Traditional neural networks, while powerful, often produce overconfident predictions, even when presented with out-of-distribution data, that are different from the training data or ambiguous inputs. This lack of calibrated uncertainty can undermine trust in AI systems.
In this work, we present recent advances in uncertainty quantification (UQ) methods for deep learning and demonstrate their practical relevance across both classification and regression tasks. These UQ methods are further applied to downstream tasks such as out-of-distribution detection, active learning, and model calibration. The results illustrate how uncertainty-aware models can more accurately reflect predictive confidence and contribute to the development of more robust and interpretable AI systems.

Authors

Aishwarya Venkataramanan (Computer Vision Group) Joachim Denzler (Computer Vision Group)

Presentation materials

There are no materials yet.