GOEN Pipeline Outperforms OOD Detection Baselines, CenterLoss Found Harmful
The recently developed machine learning pipeline, named GOEN (Geometry-Optimised Epistemic Network), sets a new benchmark in out-of-distribution (OOD) detection, achieving an impressive average OOD AUROC of 0.9483. This innovative approach integrates multi-scale features, L2 normalization, Mahalanobis distance, and a calibration head that is trained using actual challenging OOD examples. Researchers conducted systematic ablation studies and found that CenterLoss, which is intended to enhance feature compactness, inadvertently hampers OOD performance, causing a drop in AUROC to 0.9366, even though it boosts classification accuracy. The top-performing variant, GOEN-NoCenterLoss, outperforms deep ensembles (0.8827), KNN (0.8967), and ODIN (0.8870) on dataset C, highlighting the need for feature representations that address epistemic uncertainty.
Key facts
- GOEN achieves average OOD AUROC of 0.9483
- CenterLoss reduces OOD AUROC from 0.9483 to 0.9366
- GOEN-NoCenterLoss is the best variant
- Deep ensembles achieve 0.8827 AUROC
- KNN achieves 0.8967 AUROC
- ODIN achieves 0.8870 AUROC
- Pipeline includes multi-scale features, L2 normalisation, Mahalanobis distance, calibration head
- Study published on arXiv with ID 2605.21493
Entities
Institutions
- arXiv