{"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/do-deep-generative-models-know-what-they-dont","title":"Do Deep Generative Models Know What They Don't Know?","arxiv_id":"1810.09136","date":"2018-10-22","proceeding":"ICLR 2019 5","authors":["Eric Nalisnick","Akihiro Matsukawa","Yee Whye Teh","Dilan Gorur","Balaji Lakshminarayanan"],"abstract":"A neural network deployed in the wild may be asked to make predictions for\ninputs that were drawn from a different distribution than that of the training\ndata. A plethora of work has demonstrated that it is easy to find or synthesize\ninputs for which a neural network is highly confident yet wrong. Generative\nmodels are widely viewed to be robust to such mistaken confidence as modeling\nthe density of the input features can be used to detect novel,\nout-of-distribution inputs. In this paper we challenge this assumption. We find\nthat the density learned by flow-based models, VAEs, and PixelCNNs cannot\ndistinguish images of common objects such as dogs, trucks, and horses (i.e.\nCIFAR-10) from those of house numbers (i.e. SVHN), assigning a higher\nlikelihood to the latter when the model is trained on the former. Moreover, we\nfind evidence of this phenomenon when pairing several popular image data sets:\nFashionMNIST vs MNIST, CelebA vs SVHN, ImageNet vs CIFAR-10 / CIFAR-100 / SVHN.\nTo investigate this curious behavior, we focus analysis on flow-based\ngenerative models in particular since they are trained and evaluated via the\nexact marginal likelihood. We find such behavior persists even when we restrict\nthe flows to constant-volume transformations. These transformations admit some\ntheoretical analysis, and we show that the difference in likelihoods can be\nexplained by the location and variances of the data and the model curvature.\nOur results caution against using the density estimates from deep generative\nmodels to identify inputs similar to the training distribution until their\nbehavior for out-of-distribution inputs is better understood.","url_abs":"http://arxiv.org/abs/1810.09136v3","url_pdf":"http://arxiv.org/pdf/1810.09136v3.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":"do-deep-generative-models-know-what-they-dont","repo_url":"https://github.com/TaikiInoue/STAD","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":{"status":"ok"}},{"paper_slug":"do-deep-generative-models-know-what-they-dont","repo_url":"https://github.com/glouppe/info8010-deep-learning","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":null},{"paper_slug":"do-deep-generative-models-know-what-they-dont","repo_url":"https://github.com/johnpjust/UMAP","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"none","reach":{"status":"ok"}},{"paper_slug":"do-deep-generative-models-know-what-they-dont","repo_url":"https://github.com/y0ast/Glow-PyTorch","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":null}],"tasks":[],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1810.09136","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}