Search papers, labs, and topics across Lattice.
This paper systematically investigates the curvature discrepancy in loss landscapes between in-distribution (ID) and out-of-distribution (OOD) data, revealing that OOD inputs exhibit significantly larger Hessian curvature. The authors introduce Fold, a lightweight OOD detection method that utilizes feature Hessian and partial feature normalization to enhance ID-OOD separability without the computational burden of traditional curvature approximations. Experimental results demonstrate that Fold outperforms existing methods, achieving a 1.63% improvement in average AUROC and a 2.30% reduction in FPR95, all while maintaining efficiency comparable to a standard forward pass.
OOD inputs show larger Hessian curvature than ID data, leading to a novel and efficient detection method that outperforms traditional approaches.
Detecting out-of-distribution (OOD) data is crucial for reliable machine learning deployment. Among detection strategies, post-hoc methods are particularly attractive due to their efficiency, as they operate directly on pre-trained networks without requiring retraining. Within this paradigm, one promising direction exploits loss-landscape curvature to estimate model uncertainty; however, such methods incur substantial computational cost and rely on implicit assumptions about how landscape flatness differs between in-distribution (ID) and OOD data. In this work, we provide the first systematic investigation of this curvature discrepancy and show that OOD inputs exhibit larger Hessian curvature than ID data, with the gap widening under stronger distributional shifts. Motivated by these observations, we propose Fold, a lightweight flatness-modulated OOD detector that leverages the feature Hessian and partial feature normalization to improve ID-OOD separability while avoiding costly parameter-space curvature approximations. To optimally adapt this normalization across diverse datasets, we further introduce AutoFold, a self-supervised tuning scheme that synthesizes pseudo-OOD samples via ID logit masking for automatic calibration without requiring external data. Experiments on OOD benchmarks show that Fold outperforms prior methods, improving the average AUROC by 1.63% and reducing FPR95 by 2.30%, while maintaining computational efficiency comparable to a standard forward pass. Supported by theoretical analysis and extensive ablations, Fold provides a principled and practical solution for robust real-world deployment.
