level: research
a paper introduces goen, a pipeline for out-of-distribution detection that combines multi-scale features, l2 normalization, mahalanobis distance, and a calibration head trained with real hard ood examples. the method achieves an average ood auroc of 0.9483, outperforming deep ensembles, knn, and odin. the work focuses on the gap between classification accuracy and epistemic uncertainty, arguing that features optimized only for accuracy may fail at ood detection.
a key finding is that centerloss, a regularizer that makes features more compact, actually hurts ood detection. adding centerloss dropped the average ood auroc from 0.9483 to 0.9366, even though classification accuracy improved. the authors suggest that collapsing features into tight clusters removes the spread needed to distinguish in-distribution from out-of-distribution samples. the best variant, goen-nocenterloss, avoids this issue and keeps feature representations more informative for uncertainty.
the pipeline uses features from multiple layers of a network, normalizes them with l2, and computes mahalanobis distance to measure how far a sample is from the training distribution. a calibration head is trained on real hard ood examples to improve the reliability of confidence scores. systematic ablations show that each component contributes to the final performance, with multi-scale features and the absence of centerloss being particularly important.
why it matters: it shows that common tricks for improving classification can backfire on safety-critical tasks like detecting unfamiliar inputs, and offers a simple, high-performing alternative.