{"oa":"1","main_file_link":[{"url":"https://arxiv.org/pdf/2305.13764.pdf","open_access":"1"}],"external_id":{"arxiv":["2305.13764"]},"status":"public","author":[{"last_name":"Lienen","first_name":"Julian","full_name":"Lienen, Julian"},{"first_name":"Eyke","last_name":"Hüllermeier","full_name":"Hüllermeier, Eyke"}],"date_updated":"2023-06-30T14:20:31Z","title":"Mitigating Label Noise through Data Ambiguation","type":"preprint","user_id":"44040","year":"2023","_id":"45244","abstract":[{"text":"Label noise poses an important challenge in machine learning, especially in\r\ndeep learning, in which large models with high expressive power dominate the\r\nfield. Models of that kind are prone to memorizing incorrect labels, thereby\r\nharming generalization performance. Many methods have been proposed to address\r\nthis problem, including robust loss functions and more complex label correction\r\napproaches. Robust loss functions are appealing due to their simplicity, but\r\ntypically lack flexibility, while label correction usually adds substantial\r\ncomplexity to the training setup. In this paper, we suggest to address the\r\nshortcomings of both methodologies by \"ambiguating\" the target information,\r\nadding additional, complementary candidate labels in case the learner is not\r\nsufficiently convinced of the observed training label. More precisely, we\r\nleverage the framework of so-called superset learning to construct set-valued\r\ntargets based on a confidence threshold, which deliver imprecise yet more\r\nreliable beliefs about the ground-truth, effectively helping the learner to\r\nsuppress the memorization effect. In an extensive empirical evaluation, our\r\nmethod demonstrates favorable learning behavior on synthetic and real-world\r\nnoise, confirming the effectiveness in detecting and correcting erroneous\r\ntraining labels.","lang":"eng"}],"language":[{"iso":"eng"}],"citation":{"chicago":"Lienen, Julian, and Eyke Hüllermeier. “Mitigating Label Noise through Data Ambiguation.” <i>ArXiv:2305.13764</i>, 2023.","mla":"Lienen, Julian, and Eyke Hüllermeier. “Mitigating Label Noise through Data Ambiguation.” <i>ArXiv:2305.13764</i>, 2023.","ieee":"J. Lienen and E. Hüllermeier, “Mitigating Label Noise through Data Ambiguation,” <i>arXiv:2305.13764</i>. 2023.","bibtex":"@article{Lienen_Hüllermeier_2023, title={Mitigating Label Noise through Data Ambiguation}, journal={arXiv:2305.13764}, author={Lienen, Julian and Hüllermeier, Eyke}, year={2023} }","short":"J. Lienen, E. Hüllermeier, ArXiv:2305.13764 (2023).","ama":"Lienen J, Hüllermeier E. Mitigating Label Noise through Data Ambiguation. <i>arXiv:230513764</i>. Published online 2023.","apa":"Lienen, J., &#38; Hüllermeier, E. (2023). Mitigating Label Noise through Data Ambiguation. In <i>arXiv:2305.13764</i>."},"date_created":"2023-05-24T05:28:34Z","publication":"arXiv:2305.13764"}