Out-of-Distribution Detection Using Layerwise Uncertainty in Deep Neural Networks

In this paper, we tackle the problem of detecting samples that are not drawn from the training distribution, i.e., out-of-distribution (OOD) samples, in classification. Many previous studies have attempted to solve this problem by regarding samples with low classification confidence as OOD examples using deep neural networks (DNNs). However, on difficult datasets or models with low classification ability, these methods incorrectly regard in-distribution samples close to the decision boundary as OOD samples. This problem arises because their approaches use only the features close to the output layer and disregard the uncertainty of the features. Therefore, we propose a method that extracts the uncertainties of features in each layer of DNNs using a reparameterization trick and combines them. In experiments, our method outperforms the existing methods by a large margin, achieving state-of-the-art detection performance on several datasets and classification models. For example, our method increases the AUROC score of prior work (83.8%) to 99.8% in DenseNet on the CIFAR-100 and Tiny-ImageNet datasets.