Abstract

With the rapid development of AI and generative AI, the focus on quantifying uncertainty in these systems should become increasingly important. This applies even to “simple” models in “classical” supervised learning. We motivate the neccessity but also the obstacles for quantifying uncertainty for AI systems. Inducing variation and variability does not mean capturing uncertainty and the question remains, how confident can we be.
In our view the answer boils down to statistical perspectives and the role of data, data aquisation and data quality for training models and assessing their uncertainty and confidence, respectively.