Don't Collapse Your Features: Why CenterLoss Hurts OOD Detection and Multi-Scale Mahalanobis Wins

In three linesGOEN (Geometry-Optimised Epistemic Network) combines multi-scale features, L2 normalisation, Mahalanobis distance, and calibration to detect out-of-distribution inputs. Key finding: CenterLoss degrades OOD detection (AUROC 0.9366 vs 0.9483 without), despite improving classification accuracy. GOEN-NoCenterLoss achieves 0.9483 AUROC on CIFAR-10, outperforming deep ensembles (0.8827), KNN (0.8967), and ODIN (0.8870).

Summary generated by Claude — human-verified