Discriminative methods learn to map data to labels but different models with identical i.i.d. test performance may use completely different features for decision making. This can be demonstrated e.g. by the use of carefully designed architectures that exclude the use of certain features. In addition, we use generative methods such as adversarial, controversial, or style transfer stimuli that can help to reveal the features used by a neural network, or that are used during inference (analysis-by-synthesis). Sometimes these methods also facilitate aesthetically compelling image manipulations similar to artistic styles.