Out-Of-Distribution Detection With Subspace Techniques And Probabilistic Modeling Of Features

This paper presents a principled approach for detecting out-of-distribution\n(OOD) samples in deep neural networks (DNN). Modeling probability distributions\non deep features has recently emerged as an effective, yet computationally\ncheap method to detect OOD samples in DNN. However, the features produced by a\nDNN at any given layer do not fully occupy the corresponding high-dimensional\nfeature space. We apply linear statistical dimensionality reduction techniques\nand nonlinear manifold-learning techniques on the high-dimensional features in\norder to capture the true subspace spanned by the features. We hypothesize that\nsuch lower-dimensional feature embeddings can mitigate the curse of\ndimensionality, and enhance any feature-based method for more efficient and\neffective performance. In the context of uncertainty estimation and OOD, we\nshow that the log-likelihood score obtained from the distributions learnt on\nthis lower-dimensional subspace is more discriminative for OOD detection. We\nalso show that the feature reconstruction error, which is the $L_2$-norm of the\ndifference between the original feature and the pre-image of its embedding, is\nhighly effective for OOD detection and in some cases superior to the\nlog-likelihood scores. The benefits of our approach are demonstrated on image\nfeatures by detecting OOD images, using popular DNN architectures on commonly\nused image datasets such as CIFAR10, CIFAR100, and SVHN.\n