9ec51f6eb240fb631a35864e13737bca-AuthorFeedback.pdf
–Neural Information Processing Systems
The polyhedral loss given in Theorem 2 would likely37 not be "computed" per se, as the discrete loss typically depends on the number of labelsn, and one would want a38 mathematical expression for the loss in terms ofn. This expression, which is essentially the Fenchel conjugate of39 a polyhedral function, follows from standard results in convex analysis [Rockafellar, 1997, Thm 19.1, Thm 19.2].40
Neural Information Processing Systems
Feb-13-2026, 06:18:39 GMT