On the inconsistency of separable losses for structured prediction

Caio Corro

In this paper, we prove that separable negative log-likelihood losses for structured prediction are not necessarily Bayes consistent, or, in other words, minimizing these losses ma…