Uncertainty in Deep Learning
Semester
Winter term
Contents
Modern deep learning models achieve excellent performance across a wide range of tasks. However, they are also known to be overconfident when presented with novel data that was not encountered during training. This overconfidence can be particularly problematic in safety-critical applications such as healthcare, autonomous driving, and scientific discovery, where understanding a model’s confidence or uncertainty in its predictions is essential.
The objective of this course is to equip students with both the theoretical foundations and practical tools needed to quantify and reason about uncertainty in deep learning. The course begins with an introduction to the two primary types of uncertainty: epistemic (model uncertainty) and aleatoric (data uncertainty), and explains why uncertainty quantification is critical in modern AI systems. Students are also introduced to the concept of model calibration, which involves aligning predicted probabilities with actual outcome frequencies, which is a crucial first step toward building trustworthy and reliable models. The course then covers a range of methods for uncertainty estimation, including:
- Bayesian Neural Networks
- Monte Carlo Dropout
- Deep Ensembles
- Conformal Prediction
The final part of the course focuses on real-world applications, demonstrating how uncertainty can improve model robustness and decision-making in domains such as active learning, scientific machine learning, and large language models.
A prior knowledge on deep learning is highly recommended.