Don't Collapse Your Features: Why CenterLoss Hurts OOD Detection and Multi-Scale Mahalanobis Wins
Researchers have developed a new pipeline called GOEN that improves the detection of out-of-distribution inputs in machine learning systems. This method combines multi-scale features, L2 normalization, Mahalanobis distance, and a calibration head trained with real out-of-distribution examples. Their findings indicate that CenterLoss, a common regularizer for feature compactness, actually degrades out-of-distribution detection performance, while GOEN-NoCenterLoss achieved a superior OOD AUROC of 0.9483 on CIFAR-10 benchmarks. AI
IMPACT Enhances AI safety by improving the ability of models to recognize and flag unfamiliar or out-of-distribution data.