ConjNorm: A New Framework for OOD Detection Density Estimation

A novel theoretical framework for out-of-distribution (OOD) detection, named ConjNorm, has been introduced in a paper available on arXiv (2402.17888). This framework is based on Bregman divergence and broadens the scope of distribution considerations to encompass an exponential family of distributions. By implementing a conjugation constraint, the approach reinterprets the design of density functions as a quest for the ideal norm coefficient p tailored to a specific dataset. Additionally, the authors present an unbiased analytical method to tackle the normalization's computational difficulties. This research seeks to offer a cohesive viewpoint on density-based score design for OOD detection, overcoming the shortcomings of current methods that depend on logits, distances, or strict data distribution assumptions.