{"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/towards-a-definition-of-disentangled","title":"Towards a Definition of Disentangled Representations","arxiv_id":"1812.02230","date":"2018-12-05","proceeding":null,"authors":["Irina Higgins","David Amos","David Pfau","Sebastien Racaniere","Loic Matthey","Danilo Rezende","Alexander Lerchner"],"abstract":"How can intelligent agents solve a diverse set of tasks in a data-efficient\nmanner? The disentangled representation learning approach posits that such an\nagent would benefit from separating out (disentangling) the underlying\nstructure of the world into disjoint parts of its representation. However,\nthere is no generally agreed-upon definition of disentangling, not least\nbecause it is unclear how to formalise the notion of world structure beyond toy\ndatasets with a known ground truth generative process. Here we propose that a\nprincipled solution to characterising disentangled representations can be found\nby focusing on the transformation properties of the world. In particular, we\nsuggest that those transformations that change only some properties of the\nunderlying world state, while leaving all other properties invariant, are what\ngives exploitable structure to any kind of data. Similar ideas have already\nbeen successfully applied in physics, where the study of symmetry\ntransformations has revolutionised the understanding of the world structure. By\nconnecting symmetry transformations to vector representations using the\nformalism of group and representation theory we arrive at the first formal\ndefinition of disentangled representations. Our new definition is in agreement\nwith many of the current intuitions about disentangling, while also providing\nprincipled resolutions to a number of previous points of contention. While this\nwork focuses on formally defining disentangling - as opposed to solving the\nlearning problem - we believe that the shift in perspective to studying data\ntransformations can stimulate the development of better representation learning\nalgorithms.","url_abs":"http://arxiv.org/abs/1812.02230v1","url_pdf":"http://arxiv.org/pdf/1812.02230v1.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":"towards-a-definition-of-disentangled","repo_url":"https://github.com/IndustAI/learning-group-structure","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":null}],"tasks":[{"task_slug":"representation-learning","task_name":"Representation Learning"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"syntology_url":"https://syntology.ai/paper/1812.02230","atlas_url":"https://app.syntology.ai/?focus=1812.02230","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}