Measuring how much each training example contributes to a model's final performance using gradient-based methods.