The proceedings version of this paper contained an incorrect claim in Lemma 1
for outputs that can be negative.  Because the paper focused on classification,
where DNN scores are in [0,1], the error had no impact on the claims or
experimental results for classification.  The updated version of the paper,
linked on this website, (1) corrects the error by constraining Lemma 1 to
positive outputs and (2) shows how to support negative outputs (and therefore
many regression problems) in the Appendix/Lemma 2.