Current browse context:
cs.LG
Change to browse by:
References & Citations
Computer Science > Machine Learning
Title: On the inconsistency of separable losses for structured prediction
(Submitted on 25 Jan 2023)
Abstract: 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 may not result in a model that predicts the most probable structure in the data distribution for a given input. This fact opens the question of whether these losses are well-adapted for structured prediction and, if so, why.
Link back to: arXiv, form interface, contact.