GOEN beats CenterLoss: Multi-Scale Mahalanobis achieves 0.948 OOD AUROC

Rahul D Ray’s new paper challenges a core assumption in machine learning: that better classification features automatically improve out-of-distribution (OOD) detection. The author introduces GOEN (Geometry-Optimised Epistemic Network), a simple pipeline that combines multi-scale features, L2 normalization, Mahalanobis distance, and a calibration head trained with real hard OOD examples. On CIFAR-10, GOEN achieves an average OOD AUROC of 0.9483, significantly outperforming deep ensembles (0.8827), KNN (0.8967), and ODIN (0.8870).

Through systematic ablation, Ray uncovers a counter-intuitive finding: CenterLoss, a popular regularizer that pulls features into tight clusters for improved classification accuracy, actually degrades OOD detection performance. Adding CenterLoss drops the AUROC from 0.9483 to 0.9366. The root cause: overly compact feature spaces compress inter-class margins and distort the covariance structure needed for effective OOD detection. The best variant, GOEN-NoCenterLoss, avoids this pitfall while maintaining competitive in-distribution accuracy.

GOEN is efficient, training in under 20 minutes on a single GPU. The paper provides a practical blueprint for building AI systems that reliably recognize their own limitations—critical for safety-critical deployments in healthcare, autonomous driving, and finance. The results challenge the prevailing assumption that better classification geometry automatically leads to better epistemic uncertainty.