Description: "In this paper, we use influence functions — a classic technique from robust statistics — to trace a model’s prediction through the
learning algorithm and back to its training data,
thereby identifying training points most responsible for a given prediction. "