<ul data-eligibleForWebStory="true"><li>Out-of-distribution recognition is a crucial issue in deep learning to identify samples not part of the original training data.</li><li>This study suggests that effective hierarchical hyperbolic embedding is essential for distinguishing between in- and out-of-distribution samples.</li><li>Balanced Hyperbolic Learning is introduced, optimizing class embedding by balancing hierarchical distortion and subhierarchy distribution.</li><li>Hyperbolic prototypes derived from these embeddings are used for classification on in-distribution data.</li><li>Existing out-of-distribution scoring functions are adapted to work with hyperbolic prototypes in this study.</li><li>Empirical assessments involving 13 datasets and 13 scoring functions demonstrate the superiority of hyperbolic embeddings over existing out-of-distribution methods with the same training and backbone data.</li><li>Comparison with other hyperbolic models and contrastive methods also show the effectiveness of the proposed hyperbolic embeddings.</li><li>The hyperbolic embeddings additionally support hierarchical out-of-distribution generalization, providing a native advantage.</li></ul>

Balanced Hyperbolic Embeddings Are Natural Out-of-Distribution Detectors

Discover more