Out-of-Distribution Detection in Deep Learning Models: A Feature Space-Based Approach

The deployment of Deep Learning models requires attention to certain aspects not typically considered during the training phase. One of these is identifying and labeling samples from unknown classes. This is the goal of out-of-distribution (OOD) detection, which enhances the robustness of models in open-world scenarios. There are numerous methods for addressing this problem, using different feature spaces to distinguish between in-distribution and OOD data. One such method is the Open Principal Component Score (OpenPCS), a technique developed for open-set semantic segmentation using intermediate features of a fully convolutional network. This article introduces an extension of OpenPCS for multi-class classification, called OpenPCS-Class. We evaluate our approach in image and text classification tasks using various benchmark datasets and OOD detection methods. We also assess the effect of intermediate layers and network architectures on the OOD detection task. Our method outperformed other methods by up to 6.2% in terms of AUROC and 91% in terms of FPR95.