A learned binary or continuous mask that highlights which parts of an input are important for a prediction.