Uncertainty Estimation of Neural Networks
Tillman Radmer, Fabian Hüger & Nico Schmidt - 3 years ago
Neural networks are complex statistical systems that are notoriously hard to
interpret. When using neural networks in an automotive context, it is crucial to
understand when and why a model struggles to predict a scene correctly. In our
quest to continuously improve our models, uncertainty estimation is a centerpiece
in understanding what data we need to collect to improve the predictions of our AI
models. This deep dive will explain several methods that can be applied to estimate
uncertainty of neural networks and discuss with the participants the advantages,
pitfalls and lessons learned of different approaches.