The issue with machine learning models is that they can provide an incorrect diagnosis or solution with complete confidence, especially when encountering an outlier in a data set. The ability for the network to identify and communicate its uncertainty is crucial in mission-critical applications like medical diagnosis and automated driving.