{"external_id":{"arxiv":["2305.00983"]},"main_file_link":[{"open_access":"1","url":"https://arxiv.org/pdf/2305.00983.pdf"}],"oa":"1","date_created":"2023-05-05T11:37:00Z","publication":"arXiv:2305.00983","_id":"44512","year":"2023","citation":{"ama":"Uhlemeyer S, Lienen J, Hüllermeier E, Gottschalk H. Detecting Novelties with Empty Classes. <i>arXiv:230500983</i>. Published online 2023.","apa":"Uhlemeyer, S., Lienen, J., Hüllermeier, E., &#38; Gottschalk, H. (2023). Detecting Novelties with Empty Classes. In <i>arXiv:2305.00983</i>.","ieee":"S. Uhlemeyer, J. Lienen, E. Hüllermeier, and H. Gottschalk, “Detecting Novelties with Empty Classes,” <i>arXiv:2305.00983</i>. 2023.","short":"S. Uhlemeyer, J. Lienen, E. Hüllermeier, H. Gottschalk, ArXiv:2305.00983 (2023).","mla":"Uhlemeyer, Svenja, et al. “Detecting Novelties with Empty Classes.” <i>ArXiv:2305.00983</i>, 2023.","chicago":"Uhlemeyer, Svenja, Julian Lienen, Eyke Hüllermeier, and Hanno Gottschalk. “Detecting Novelties with Empty Classes.” <i>ArXiv:2305.00983</i>, 2023.","bibtex":"@article{Uhlemeyer_Lienen_Hüllermeier_Gottschalk_2023, title={Detecting Novelties with Empty Classes}, journal={arXiv:2305.00983}, author={Uhlemeyer, Svenja and Lienen, Julian and Hüllermeier, Eyke and Gottschalk, Hanno}, year={2023} }"},"abstract":[{"text":"For open world applications, deep neural networks (DNNs) need to be aware of\r\npreviously unseen data and adaptable to evolving environments. Furthermore, it\r\nis desirable to detect and learn novel classes which are not included in the\r\nDNNs underlying set of semantic classes in an unsupervised fashion. The method\r\nproposed in this article builds upon anomaly detection to retrieve\r\nout-of-distribution (OoD) data as candidates for new classes. We thereafter\r\nextend the DNN by $k$ empty classes and fine-tune it on the OoD data samples.\r\nTo this end, we introduce two loss functions, which 1) entice the DNN to assign\r\nOoD samples to the empty classes and 2) to minimize the inner-class feature\r\ndistances between them. Thus, instead of ground truth which contains labels for\r\nthe different novel classes, the DNN obtains a single OoD label together with a\r\ndistance matrix, which is computed in advance. We perform several experiments\r\nfor image classification and semantic segmentation, which demonstrate that a\r\nDNN can extend its own semantic space by multiple classes without having access\r\nto ground truth.","lang":"eng"}],"language":[{"iso":"eng"}],"date_updated":"2023-05-05T11:39:10Z","title":"Detecting Novelties with Empty Classes","user_id":"44040","type":"preprint","author":[{"first_name":"Svenja","last_name":"Uhlemeyer","full_name":"Uhlemeyer, Svenja"},{"last_name":"Lienen","first_name":"Julian","id":"44040","full_name":"Lienen, Julian"},{"full_name":"Hüllermeier, Eyke","id":"48129","first_name":"Eyke","last_name":"Hüllermeier"},{"last_name":"Gottschalk","first_name":"Hanno","full_name":"Gottschalk, Hanno"}],"status":"public"}