Out-of-Distribution Detection with Prototype Similarity

Despite their widely known success in various applications, deep models have difficulty distinguishing the distributional uncertainties. This challenge, known as out-of-distribution sample detection, is critical in high-stakes applications such as self-driving vehicles and medical imaging. Various studies focus on frameworks based on the network’s activations at logit and penultimate layers to detect out-of-distribution samples. However, the potential benefit of using intermediate layer activations for out-of-distribution detection is yet to be thoroughly investigated. In this paper, we propose an out-of-distribution detection framework that employs the activations of several layers, including the intermediate layers of a ResNet model. The proposed approach does not modify the complex backbone architecture training but utilizes its pre-trained activations to train prototypes. Finally, an out-of-distribution classifier is trained with the similarity scores extracted from prototypes using both in and out-of-distribution samples. We showed that training prototypes to obtain scores improves the detection performance. Our results on the various benchmark datasets are on par with the state-of-the-art. Implementation of the proposed framework can be found at: https://github.com/ozanzgur/prototype-similarity/tree/main