Density-Ratio Losses for Post-Hoc Learning to Defer
Researchers have introduced a novel post-hoc Learning to Defer (L2D) framework that reframes the problem through the lens of ideal distributions. This approach defines deferral by calculating the density-ratio between a model's and an expert's ideal distributions. The derived DR CPE losses allow for adjustable deferral rates without the need for retraining, and experimental results show competitive performance and robustness across various datasets. AI
IMPACT Introduces a new theoretical framework for model deferral, potentially improving system reliability and interpretability.