NeurIPS 2020

Learning by Minimizing the Sum of Ranked Range


Meta Review

The paper introduces a novel aggregate loss function for binary and multiclass/multilabel classification to obtain models being robust against outliers. The reviewers are, in general, positive about the paper, however, there are some flaws making the paper a borderline case. The paper should discuss in depth the relation to the classical approaches and perhaps use them in the experimental studies. The comparison with the maximum loss seems to be slightly unfair as the authors use its vanilla variant. The complexity of the approach is not clearly discussed.