{"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/unifying-heterogeneous-classifiers-with","title":"Unifying Heterogeneous Classifiers with Distillation","arxiv_id":"1904.06062","date":"2019-04-12","proceeding":"CVPR 2019 6","authors":["Jayakorn Vongkulbhisal","Phongtharin Vinayavekhin","Marco Visentini-Scarzanella"],"abstract":"In this paper, we study the problem of unifying knowledge from a set of\nclassifiers with different architectures and target classes into a single\nclassifier, given only a generic set of unlabelled data. We call this problem\nUnifying Heterogeneous Classifiers (UHC). This problem is motivated by\nscenarios where data is collected from multiple sources, but the sources cannot\nshare their data, e.g., due to privacy concerns, and only privately trained\nmodels can be shared. In addition, each source may not be able to gather data\nto train all classes due to data availability at each source, and may not be\nable to train the same classification model due to different computational\nresources. To tackle this problem, we propose a generalisation of knowledge\ndistillation to merge HCs. We derive a probabilistic relation between the\noutputs of HCs and the probability over all classes. Based on this relation, we\npropose two classes of methods based on cross-entropy minimisation and matrix\nfactorisation, which allow us to estimate soft labels over all classes from\nunlabelled samples and use them in lieu of ground truth labels to train a\nunified classifier. Our extensive experiments on ImageNet, LSUN, and Places365\ndatasets show that our approaches significantly outperform a naive extension of\ndistillation and can achieve almost the same accuracy as classifiers that are\ntrained in a centralised, supervised manner.","url_abs":"http://arxiv.org/abs/1904.06062v1","url_pdf":"http://arxiv.org/pdf/1904.06062v1.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":"unifying-heterogeneous-classifiers-with","repo_url":"https://github.com/aitorlucasc/uhc_distillation","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":0,"framework":"pytorch","reach":{"status":"ok"}}],"tasks":[{"task_slug":"knowledge-distillation","task_name":"Knowledge Distillation"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1904.06062","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}