84899ae725ba49884f4c85c086f1b340-AuthorFeedback.pdf

Theorem8givesaproper2 form of generalization bound: with high probability, achieving a small value of the empirical surrogateˆL(q,D)3 guaranteesaboundonthe truesuboptimalitygap Gτ(f q). The connection between smoothed risk and the KL8 divergence (8) isknown. We appreciate the criticism of the experimental exposition, and will16 improve it as suggested, including adding learning curves and standard deviations. Below are the more detailed17 experimental results(interms oftestmisclassification) for CIFAR10, inboth the fully observed( 2.2) and partially18 observed( 3.6)cases. F was actually defined in Line 61, and its use in Proposition 2 was consistent with this definition and subsequent35 occurrences.