{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/learning-with-biased-complementary-labels","title":"Learning with Biased Complementary Labels","arxiv_id":"1711.09535","date":"2017-11-27","proceeding":"ECCV 2018 9","authors":["Xiyu Yu","Tongliang Liu","Mingming Gong","DaCheng Tao"],"abstract":"In this paper, we study the classification problem in which we have access to\neasily obtainable surrogate for true labels, namely complementary labels, which\nspecify classes that observations do \\textbf{not} belong to. Let $Y$ and\n$\\bar{Y}$ be the true and complementary labels, respectively. We first model\nthe annotation of complementary labels via transition probabilities\n$P(\\bar{Y}=i|Y=j), i\\neq j\\in\\{1,\\cdots,c\\}$, where $c$ is the number of\nclasses. Previous methods implicitly assume that $P(\\bar{Y}=i|Y=j), \\forall\ni\\neq j$, are identical, which is not true in practice because humans are\nbiased toward their own experience. For example, as shown in Figure 1, if an\nannotator is more familiar with monkeys than prairie dogs when providing\ncomplementary labels for meerkats, she is more likely to employ \"monkey\" as a\ncomplementary label. We therefore reason that the transition probabilities will\nbe different. In this paper, we propose a framework that contributes three main\ninnovations to learning with \\textbf{biased} complementary labels: (1) It\nestimates transition probabilities with no bias. (2) It provides a general\nmethod to modify traditional loss functions and extends standard deep neural\nnetwork classifiers to learn with biased complementary labels. (3) It\ntheoretically ensures that the classifier learned with complementary labels\nconverges to the optimal one learned with true labels. Comprehensive\nexperiments on several benchmark datasets validate the superiority of our\nmethod to current state-of-the-art methods.","url_abs":"http://arxiv.org/abs/1711.09535v3","url_pdf":"http://arxiv.org/pdf/1711.09535v3.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"learning-with-biased-complementary-labels","repo_url":"https://github.com/takashiishida/comp","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":{"status":"ok","spdx":"MIT"}}],"tasks":[],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1711.09535","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}