[2G4-GS-2f] 機械学習：データ分布

[2G4-GS-2f-04] A Study on Detecting Out-of-Distribution based on Generative Models Trained for Each Discriminant Class

〇Ryota Matsunae¹, Fuyu Saito¹, Haruka Yamashita², Masayuki Goto¹ (1. Waseda University, 2. Sophia University)

Keywords:out-of-distribution Detection, unsupervised learning, generative model, normalizing flow, deep learning

A discriminative model may classify input that do not belong to any of the discriminant classes with high confidence. Such data is called out-of-distribution (OOD) data, while the target data of the discriminative model is called in-distribution data. In order to detect OOD data, a method using the likelihood ratio has been proposed. In this method, the likelihood ratio calculated by two generative models is used as the detection index. The generative model used for OOD detection is required to estimate the true distribution of in-distribution data accurately. On the other hand, in-distribution data may follow different distributions for each class, and the distribution structure of each class is assumed to be simpler than that of all classes together. In this study, we propose OOD detection method using generative models trained for each class. We also conduct experiments using image datasets to show the effectiveness of the proposed method.