{"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/training-deep-neural-networks-on-noisy-labels","title":"Training Deep Neural Networks on Noisy Labels with Bootstrapping","arxiv_id":"1412.6596","date":"2014-12-20","proceeding":null,"authors":["Scott Reed","Honglak Lee","Dragomir Anguelov","Christian Szegedy","Dumitru Erhan","Andrew Rabinovich"],"abstract":"Current state-of-the-art deep learning systems for visual object recognition\nand detection use purely supervised training with regularization such as\ndropout to avoid overfitting. The performance depends critically on the amount\nof labeled examples, and in current practice the labels are assumed to be\nunambiguous and accurate. However, this assumption often does not hold; e.g. in\nrecognition, class labels may be missing; in detection, objects in the image\nmay not be localized; and in general, the labeling may be subjective. In this\nwork we propose a generic way to handle noisy and incomplete labeling by\naugmenting the prediction objective with a notion of consistency. We consider a\nprediction consistent if the same prediction is made given similar percepts,\nwhere the notion of similarity is between deep network features computed from\nthe input data. In experiments we demonstrate that our approach yields\nsubstantial robustness to label noise on several datasets. On MNIST handwritten\ndigits, we show that our model is robust to label corruption. On the Toronto\nFace Database, we show that our model handles well the case of subjective\nlabels in emotion recognition, achieving state-of-the- art results, and can\nalso benefit from unlabeled face images with no modification to our method. On\nthe ILSVRC2014 detection challenge data, we show that our approach extends to\nvery deep networks, high resolution images and structured outputs, and results\nin improved scalable detection.","url_abs":"http://arxiv.org/abs/1412.6596v3","url_pdf":"http://arxiv.org/pdf/1412.6596v3.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":"training-deep-neural-networks-on-noisy-labels","repo_url":"https://bitbucket.org/woalsdnd/codes_and_data","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"tf","reach":null},{"paper_slug":"training-deep-neural-networks-on-noisy-labels","repo_url":"https://github.com/dr-darryl-wright/Noisy-Labels-with-Bootstrapping","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"none","reach":{"status":"ok","spdx":"MIT"}},{"paper_slug":"training-deep-neural-networks-on-noisy-labels","repo_url":"https://github.com/vfdev-5/BootstrappingLoss","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":{"status":"ok","spdx":"MIT"}}],"tasks":[{"task_slug":"emotion-recognition","task_name":"Emotion Recognition"},{"task_slug":"object-recognition","task_name":"Object Recognition"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"syntology_url":null,"atlas_url":"https://app.syntology.ai/?focus=1412.6596","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}