CenterLoss hurts OOD detection — GOEN uses multi-scale Mahalanobis instead
arXiv paper shows that CenterLoss, widely used to compact learned features, paradoxically degrades out-of-distribution (OOD) detection performance by 1.2 percentage points in AUROC despite improving classification accuracy.
GOEN (Geometry-Optimised Epistemic Network) replaces CenterLoss with multi-scale features, L2 normalisation, Mahalanobis distance, and a calibration head trained on real hard OOD examples to recover and exceed baseline performance.
The finding highlights that feature regularization optimized for classification alone conflicts with the geometric requirements of uncertainty quantification.