{"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/conditional-variance-penalties-and-domain","title":"Conditional Variance Penalties and Domain Shift Robustness","arxiv_id":"1710.11469","date":"2017-10-31","proceeding":null,"authors":["Christina Heinze-Deml","Nicolai Meinshausen"],"abstract":"When training a deep neural network for image classification, one can broadly\ndistinguish between two types of latent features of images that will drive the\nclassification. We can divide latent features into (i) \"core\" or \"conditionally\ninvariant\" features $X^\\text{core}$ whose distribution $X^\\text{core}\\vert Y$,\nconditional on the class $Y$, does not change substantially across domains and\n(ii) \"style\" features $X^{\\text{style}}$ whose distribution $X^{\\text{style}}\n\\vert Y$ can change substantially across domains. Examples for style features\ninclude position, rotation, image quality or brightness but also more complex\nones like hair color, image quality or posture for images of persons. Our goal\nis to minimize a loss that is robust under changes in the distribution of these\nstyle features. In contrast to previous work, we assume that the domain itself\nis not observed and hence a latent variable.\n  We do assume that we can sometimes observe a typically discrete identifier or\n\"$\\mathrm{ID}$ variable\". In some applications we know, for example, that two\nimages show the same person, and $\\mathrm{ID}$ then refers to the identity of\nthe person. The proposed method requires only a small fraction of images to\nhave $\\mathrm{ID}$ information. We group observations if they share the same\nclass and identifier $(Y,\\mathrm{ID})=(y,\\mathrm{id})$ and penalize the\nconditional variance of the prediction or the loss if we condition on\n$(Y,\\mathrm{ID})$. Using a causal framework, this conditional variance\nregularization (CoRe) is shown to protect asymptotically against shifts in the\ndistribution of the style variables. Empirically, we show that the CoRe penalty\nimproves predictive accuracy substantially in settings where domain changes\noccur in terms of image quality, brightness and color while we also look at\nmore complex changes such as changes in movement and posture.","url_abs":"http://arxiv.org/abs/1710.11469v5","url_pdf":"http://arxiv.org/pdf/1710.11469v5.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":"conditional-variance-penalties-and-domain","repo_url":"https://github.com/christinaheinze/core","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":1,"framework":"tf","reach":{"status":"ok"}}],"tasks":[{"task_slug":"classification","task_name":"General Classification"},{"task_slug":"image-classification","task_name":"Image Classification"},{"task_slug":"image-classification","task_name":"image-classification"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1710.11469","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}