A Study on Out-of-Distribution Detection based on Generative Models Trained for Each Discriminant Class

Abstract

Since discriminative models are usually constructed by learning a set of given training data, it is impossible to guarantee the predictive performance of data that are not generated from the same distribution as the training data. Such data are known as out-of-distribution (OOD) data, while data that follow the same distribution as the training data are referred to as in-distribution data. For practical applications, it is important to detect the OOD data before they are input into highly qualified classifiers. Recently, a likelihood ratio-based method of OOD detection has been proposed. In this method, the likelihood ratio calculated using two generative models with different noise conditions functions as a detection index to evaluate semantic information only, ignoring the background information associated with all classes of data. Here, the generative model used for OOD detection should estimate the true distribution of in-distribution data accurately; all classes are estimated together, using the conventional method. However, in-distribution data may follow a simpler distribution if each class is estimated separately; these simple distribution structures are easier to learn. For this reason, estimation accuracy can be improved by models that estimate the distribution structure of each class independently. In this study, we propose an OOD detection method that uses generative models trained independently for each class. We also conduct evaluation experiments, using image datasets to demonstrate the effectiveness of the proposed method.