{"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/born-again-neural-networks","title":"Born Again Neural Networks","arxiv_id":"1805.04770","date":"2018-05-12","proceeding":"ICML 2018 7","authors":["Tommaso Furlanello","Zachary C. Lipton","Michael Tschannen","Laurent Itti","Anima Anandkumar"],"abstract":"Knowledge distillation (KD) consists of transferring knowledge from one\nmachine learning model (the teacher}) to another (the student). Commonly, the\nteacher is a high-capacity model with formidable performance, while the student\nis more compact. By transferring knowledge, one hopes to benefit from the\nstudent's compactness. %we desire a compact model with performance close to the\nteacher's. We study KD from a new perspective: rather than compressing models,\nwe train students parameterized identically to their teachers. Surprisingly,\nthese {Born-Again Networks (BANs), outperform their teachers significantly,\nboth on computer vision and language modeling tasks. Our experiments with BANs\nbased on DenseNets demonstrate state-of-the-art performance on the CIFAR-10\n(3.5%) and CIFAR-100 (15.5%) datasets, by validation error. Additional\nexperiments explore two distillation objectives: (i) Confidence-Weighted by\nTeacher Max (CWTM) and (ii) Dark Knowledge with Permuted Predictions (DKPP).\nBoth methods elucidate the essential components of KD, demonstrating a role of\nthe teacher outputs on both predicted and non-predicted classes. We present\nexperiments with students of various capacities, focusing on the under-explored\ncase where students overpower teachers. Our experiments show significant\nadvantages from transferring knowledge between DenseNets and ResNets in either\ndirection.","url_abs":"http://arxiv.org/abs/1805.04770v2","url_pdf":"http://arxiv.org/pdf/1805.04770v2.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":"born-again-neural-networks","repo_url":"https://github.com/jincan333/lot","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":0,"framework":"pytorch","reach":null},{"paper_slug":"born-again-neural-networks","repo_url":"https://github.com/nocotan/born_again_neuralnet","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":0,"framework":"pytorch","reach":{"status":"ok","spdx":"MIT"}}],"tasks":[{"task_slug":"image-classification","task_name":"Image Classification"},{"task_slug":"knowledge-distillation","task_name":"Knowledge Distillation"},{"task_slug":"language-modeling","task_name":"Language Modeling"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1805.04770","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}