Is It All a Cluster Game? - Exploring Out-of-Distribution Detection Based on Clustering in the Embedding Space

Abstract

It is essential for safety-critical applications of deep neural networks to\ndetermine when new inputs are significantly different from the training\ndistribution. In this paper, we explore this out-of-distribution (OOD)\ndetection problem for image classification using clusters of semantically\nsimilar embeddings of the training data and exploit the differences in distance\nrelationships to these clusters between in- and out-of-distribution data. We\nstudy the structure and separation of clusters in the embedding space and find\nthat supervised contrastive learning leads to well-separated clusters while its\nself-supervised counterpart fails to do so. In our extensive analysis of\ndifferent training methods, clustering strategies, distance metrics, and\nthresholding approaches, we observe that there is no clear winner. The optimal\napproach depends on the model architecture and selected datasets for in- and\nout-of-distribution. While we could reproduce the outstanding results for\ncontrastive training on CIFAR-10 as in-distribution data, we find standard\ncross-entropy paired with cosine similarity outperforms all contrastive\ntraining methods when training on CIFAR-100 instead. Cross-entropy provides\ncompetitive results as compared to expensive contrastive training methods.\n