Uncertainty with deep learning: a practical view on out of distribution detection
Daniel Siegismund, Stephan Heyse, Stephan Steigele · 2020
Discriminative neural networks eventually fail when confronted with data that was not part of the distribution of the training data samples. Moreover, out of distribution (OOD) data points are frequently ill-predicted, which can lead to misleading results and interpretation, especially in real-world applications of such decision systems. Recently, several methods had been developed to tackle this problem. These studies often rely on quality parameters, deduced by theoretical considerations, rather than real world challenges. In this work, we explain practical considerations and test five methods for OOD detection by comparing the achievable result quality and computational cost of obtaining these results.