A Unified Decision Rule for Generalized Out-of-Distribution Detection

Summary

This paper proposes a generalized Benjamini-Hochberg (g-BH) procedure as a decision rule for out-of-distribution (OOD) detection, framing the task as hypothesis testing rather than designing a new score function. The authors report that g-BH controls the false discovery rate at a pre-specified level without assuming dependence among p-values. They also derive bounds on the expected false positive rate based on the tailed generalized Gaussian family, and experiments on several generalized OOD benchmarks show g-BH outperforms traditional threshold-based decision rules.