Sneakoscope: Revisiting Unsupervised Out-of-Distribution Detection

The problem of detecting out-of-distribution (OOD) examples in neural networks has been widely studied in the literature, with state-of-the-art techniques being supervised in that they require fine-tuning on OOD data to achieve high-quality OOD detection. But supervised OOD detection methods also have a disadvantage in that they require expensive training on OOD data, curating the OOD dataset so that it is distinguishable from the in-distribution data, and significant hyper-parameter tuning. In this work, we propose a unified evaluation suite, Sneakoscope, to revisit the problem with in-depth exploration of unsupervised OOD detection. Our surprising discovery shows that (1) model architectures play a significant role in unsupervised OOD detection performance; (2) unsupervised approaches applied on large-scale pre-trained models can achieve competitive performance compared to their supervised counterparts; and (3) unsupervised OOD detection based on Mahalanobis Distance with the support of a pre-trained model consistently outperforms other unsupervised methods by a large margin and compares favorably with results from state-of-the-art supervised OOD detection methods reported in the literature. We thus provide new baselines for unsupervised OOD detection methods.