Fig. 2

Corner cases of an ML model that result from the model itself or the model architecture are currently not considered much. However, to use an ML model in a critical application, the corner cases that cause incorrect behavior are needed for training and validation. This graphic shows a classification problem with two classes (speed limit 30 and no passing sign) highlighted by a blue and yellow area and divided by a decision boundary (dashed line). Besides the abstract traffic sign symbol, we have added some real samples of German and Scandinavian traffic signs [36, 37] to have a better impression. Some are common samples, and a few show different ML “corner cases” that could appear in the data. However, the question arises, which samples are a corner case from the point of view of the ML model and deserve the term corner case