{"intvolume":"        97","volume":97,"year":"2019","page":"4114-4124","author":[{"full_name":"Locatello, Francesco","id":"26cfd52f-2483-11ee-8040-88983bcc06d4","last_name":"Locatello","first_name":"Francesco","orcid":"0000-0002-4850-0683"},{"first_name":"Stefan","last_name":"Bauer","full_name":"Bauer, Stefan"},{"full_name":"Lucic, Mario","last_name":"Lucic","first_name":"Mario"},{"full_name":"Rätsch, Gunnar","last_name":"Rätsch","first_name":"Gunnar"},{"full_name":"Gelly, Sylvain","last_name":"Gelly","first_name":"Sylvain"},{"first_name":"Bernhard","last_name":"Schölkopf","full_name":"Schölkopf, Bernhard"},{"full_name":"Bachem, Olivier","first_name":"Olivier","last_name":"Bachem"}],"publication":"Proceedings of the 36th International Conference on Machine Learning","external_id":{"arxiv":["1811.12359"]},"citation":{"mla":"Locatello, Francesco, et al. “Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representations.” <i>Proceedings of the 36th International Conference on Machine Learning</i>, vol. 97, ML Research Press, 2019, pp. 4114–24.","ieee":"F. Locatello <i>et al.</i>, “Challenging common assumptions in the unsupervised learning of disentangled representations,” in <i>Proceedings of the 36th International Conference on Machine Learning</i>, Long Beach, CA, United States, 2019, vol. 97, pp. 4114–4124.","ama":"Locatello F, Bauer S, Lucic M, et al. Challenging common assumptions in the unsupervised learning of disentangled representations. In: <i>Proceedings of the 36th International Conference on Machine Learning</i>. Vol 97. ML Research Press; 2019:4114-4124.","ista":"Locatello F, Bauer S, Lucic M, Rätsch G, Gelly S, Schölkopf B, Bachem O. 2019. Challenging common assumptions in the unsupervised learning of disentangled representations. Proceedings of the 36th International Conference on Machine Learning. International Conference on Machine Learning vol. 97, 4114–4124.","apa":"Locatello, F., Bauer, S., Lucic, M., Rätsch, G., Gelly, S., Schölkopf, B., &#38; Bachem, O. (2019). Challenging common assumptions in the unsupervised learning of disentangled representations. In <i>Proceedings of the 36th International Conference on Machine Learning</i> (Vol. 97, pp. 4114–4124). Long Beach, CA, United States: ML Research Press.","chicago":"Locatello, Francesco, Stefan Bauer, Mario Lucic, Gunnar Rätsch, Sylvain Gelly, Bernhard Schölkopf, and Olivier Bachem. “Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representations.” In <i>Proceedings of the 36th International Conference on Machine Learning</i>, 97:4114–24. ML Research Press, 2019.","short":"F. Locatello, S. Bauer, M. Lucic, G. Rätsch, S. Gelly, B. Schölkopf, O. Bachem, in:, Proceedings of the 36th International Conference on Machine Learning, ML Research Press, 2019, pp. 4114–4124."},"conference":{"name":"International Conference on Machine Learning","start_date":"2019-06-10","location":"Long Beach, CA, United States","end_date":"2019-06-15"},"language":[{"iso":"eng"}],"_id":"14200","oa":1,"oa_version":"Preprint","date_created":"2023-08-22T14:13:08Z","user_id":"2DF688A6-F248-11E8-B48F-1D18A9856A87","publisher":"ML Research Press","scopus_import":"1","quality_controlled":"1","title":"Challenging common assumptions in the unsupervised learning of disentangled representations","status":"public","type":"conference","day":"09","month":"06","article_processing_charge":"No","date_updated":"2023-09-13T07:45:30Z","publication_status":"published","extern":"1","date_published":"2019-06-09T00:00:00Z","department":[{"_id":"FrLo"}],"abstract":[{"text":"The key idea behind the unsupervised learning of disentangled representations\r\nis that real-world data is generated by a few explanatory factors of variation\r\nwhich can be recovered by unsupervised learning algorithms. In this paper, we\r\nprovide a sober look at recent progress in the field and challenge some common\r\nassumptions. We first theoretically show that the unsupervised learning of\r\ndisentangled representations is fundamentally impossible without inductive\r\nbiases on both the models and the data. Then, we train more than 12000 models\r\ncovering most prominent methods and evaluation metrics in a reproducible\r\nlarge-scale experimental study on seven different data sets. We observe that\r\nwhile the different methods successfully enforce properties ``encouraged'' by\r\nthe corresponding losses, well-disentangled models seemingly cannot be\r\nidentified without supervision. Furthermore, increased disentanglement does not\r\nseem to lead to a decreased sample complexity of learning for downstream tasks.\r\nOur results suggest that future work on disentanglement learning should be\r\nexplicit about the role of inductive biases and (implicit) supervision,\r\ninvestigate concrete benefits of enforcing disentanglement of the learned\r\nrepresentations, and consider a reproducible experimental setup covering\r\nseveral data sets.","lang":"eng"}],"main_file_link":[{"open_access":"1","url":"https://arxiv.org/abs/1811.12359"}]}