Fig. 1. Is a model unfair due to shortcutting?

a Examples of correct and incorrect predictions that may be influenced by age shortcut learning in a chest X-ray application detecting Effusion. In this example, the prediction is incorrect for a patient with atypical age, raising the possibility of shortcut learning. b Simplified diagram illustrating how shortcutting may occur. In this schematic, the presence or absence of a particular condition y will produce changes in the image x; we therefore wish to train a model that can predict y, given x (blue arrow). However, an attribute a, such as age, may alter both the risk of developing a given condition, as well as the image. This need not be a directly direct relationship (dotted arrows). A model may learn to predict the presence of a condition by using the attribute (red arrow). When these correlations are not considerably beneficial for model performance, we consider that this path is (at least partly) a shortcut.