ReweightOOD: Loss Reweighting for Distance-based OOD Detection

Sudarshan Regmi, Bibek Panthi, Yifei Ming, Prashnna K Gyawali, Danail Stoyanov, Binod Bhattarai; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2024, pp. 131-141

Out-of-Distribution (OOD) detection is crucial for ensuring safety and reliability of neural networks in critical applications. Distance-based OOD detection is based on the assumption that OOD samples are mapped far from In-Distribution (ID) clusters in embedding space. A recent approach for obtaining OOD-detection-friendly embedding space has been contrastive optimization of pulling similar pairs and pushing apart dissimilar pairs. It assigns equal significance to all similarity instances with the implicit objective of maximizing the mean proximity between samples with their corresponding hypothetical class centroids. However the emphasis should be directed towards reducing the Minimum Enclosing Sphere (MES) for each class and achieving higher inter-class dispersion to effectively mitigate the potential for ID-OOD overlap. Optimizing low-signal dissimilar pairs might potentially act against achieving maximal inter-class dispersion while less-optimized similar pairs prevent achieving smaller MES. Based on this we propose a reweighting scheme ReweightOOD that adopts the similarity optimization which prioritizes the optimization of less-optimized contrasting pairs while assigning lower importance to already well-optimized contrasting pairs. Such a reweighting scheme serves to minimize the MES for each class while achieving maximal inter-class dispersion. Experimental results on a challenging CIFAR100 benchmark using ResNet-18 network demonstrate that ReweightOOD outperforms supervised contrastive loss by a whopping 38% in the average FPR metric. In various classification datasets our method provides a promising solution for enhancing OOD detection capabilities in neural networks.