Topological Structure Learning for Weakly-Supervised Out-of-Distribution Detection

However, in many real applications, the assumption cannot be satisfied due to the existence of unknowns. A reliable classification model ought to own the ability to say "I do not know" to out-of-distribution (OOD) data that the model has not seen before, which is the key to deploying models safely in the real world [34, 45]. For example, a wildlife monitoring system with the ability to detect OOD data will not confidently regard unknown animal categories as known categories, which is essential to help humans discover new species [35]. In medical image recognition [12, 41], models with the ability to detect OOD data can help doctors discover rare and novel diseases and prevent patients missing the best treatment period. In autonomous driving, OOD detection enables cars to evoke human control of driving in an emergency or unknown scenarios [17, 36], which contributes to safer and more reliable autonomous driving. OOD detection has received much attention because of its significance, and plenty of methods have emerged. The existing OOD detection methods can be divided into two main categories: classification-based OOD detection methods and density-based OOD detection methods. The classification-based methods contain post-hoc based methods [18, 27] and fine-tuning based methods [19, 30].