Feature-attribution methods are used to determine if a neural network is working correctly when completing a task like image classification. MIT researchers developed a way to evaluate whether these feature-attribution methods are correctly identifying the features of an image that are important to a neural network’s prediction.