Predictions of opaque black-box systems are frequently deployed in high-stakes applications such as healthcare. For such applications, it is crucial to assess how models handle samples beyond the domain of training data. While several metrics and tests exist to detect out-of-distribution (OoD) data from in-distribution (InD) data to a deep neural network (DNN), their performance varies significantly across datasets, models, and tasks, which limits their practical use. In this paper, we propose a hypothesis-driven approach to quantify whether a new sample is InD or OoD. Given a trained DNN and some input, we first feed the input through the DNN and compute an ensemble of OoD metrics, which we term latent responses. We then formulate the OoD detection problem as a hypothesis test between latent responses of different groups, and use permutation-based resampling to infer the significance of the observed latent responses under a null hypothesis. We adapt our method to detect an unseen sample of bacteria to a trained deep learning model, and show that it reveals interpretable differences between InD and OoD latent responses. Our work has implications for systematic novelty detection and informed decision-making from classifiers trained on a subset of labels.
View on arXiv