{"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/can-vaes-generate-novel-examples","title":"Can VAEs Generate Novel Examples?","arxiv_id":"1812.09624","date":"2018-12-22","proceeding":null,"authors":["Alican Bozkurt","Babak Esmaeili","Dana H. Brooks","Jennifer G. Dy","Jan-Willem van de Meent"],"abstract":"An implicit goal in works on deep generative models is that such models\nshould be able to generate novel examples that were not previously seen in the\ntraining data. In this paper, we investigate to what extent this property holds\nfor widely employed variational autoencoder (VAE) architectures. VAEs maximize\na lower bound on the log marginal likelihood, which implies that they will in\nprinciple overfit the training data when provided with a sufficiently\nexpressive decoder. In the limit of an infinite capacity decoder, the optimal\ngenerative model is a uniform mixture over the training data. More generally,\nan optimal decoder should output a weighted average over the examples in the\ntraining data, where the magnitude of the weights is determined by the\nproximity in the latent space. This leads to the hypothesis that, for a\nsufficiently high capacity encoder and decoder, the VAE decoder will perform\nnearest-neighbor matching according to the coordinates in the latent space. To\ntest this hypothesis, we investigate generalization on the MNIST dataset. We\nconsider both generalization to new examples of previously seen classes, and\ngeneralization to the classes that were withheld from the training set. In both\ncases, we find that reconstructions are closely approximated by nearest\nneighbors for higher-dimensional parameterizations. When generalizing to unseen\nclasses however, lower-dimensional parameterizations offer a clear advantage.","url_abs":"http://arxiv.org/abs/1812.09624v1","url_pdf":"http://arxiv.org/pdf/1812.09624v1.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":"can-vaes-generate-novel-examples","repo_url":"https://github.com/alicanb/vae_novel_examples","is_official":1,"mentioned_in_paper":0,"mentioned_in_github":0,"framework":"pytorch","reach":null}],"tasks":[{"task_slug":"decoder","task_name":"Decoder"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":null,"mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}