Local Uncertainty Smoothing For Few-shot OOD Detection

Pinxuan Li, Bing Cao, Changqing Zhang, Qinghua Hu

Keywords: Few-shot OOD Detection, Local Uncertainty, Label Smooth, Generalization Error Bound, Soft label

Abstract: Few-shot out-of-distribution (OOD) detection has become a critical research direction for the practical deployment of machine learning systems. Existing approaches commonly rely on auxiliary outlier data derived from in-distribution (ID) samples, such as using local image patches from training data to simulate OOD features. However, these artificially constructed OOD samples differ substantially from real OOD instances, leading to unstable learning when trained with hard OOD labels. To address this challenge, we propose a Local Uncertainty Smoothing (LUS) framework for few-shot OOD detection. Our method incorporates label smoothing and local uncertainty measure to facilitate a smooth transition between the reference distribution of local image categories, based on a general knowledge model and the target OOD distribution. This approach ensures strong OOD detection performance while preserving the model’s ability to capture detailed local-level semantic features. Furthermore, we theoretically analyze the relevance of local uncertainty from the perspective of a generalization error bound (GEB). This reveals a concrete relationship between our local uncertainty measure and the KL divergence observed during training. Accordingly, we propose a patch-wise local uncertainty to effectively identify suitable soft labels for the model throughout the learning process, achieving superior OOD detection performance. Extensive experiments on real-world OOD benchmarks validate the effectiveness of our approach. Code will be made publicly available.

Primary Area: learning theory

Submission Number: 1783