The paper addresses the problem of estimating the influence of a data point, for models trained with SGD, for non-convex losses. It then describes a data cleaning method based on this estimate. Reviewers found the theory solid, and the proposed data-cleaning method promising.