Team

Violeta Teodora Trifunov, Maha Shadaydeh, Paul Bodesheim

Motivation

In many important learning tasks, training examples are often available for only one class or a limited set of classes. Learning in this scenario is difficult, since training examples from “outside” are not available at all. A general term often used is out-of-distribution detection. For binary classification tasks, this problem is also known as one-class classification (OCC), outlier detection, or anomaly detection. In a multi-class setup, it is referred to as novelty detection,  novel object class detection, object discovery, open-set recognition, or open-world recognition. Our work in this area addresses the use of methods for deriving a so-called novelty score or anomaly score to identify samples that do not belong to the distribution of known concepts derived from given training data.