{"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/a-bayesian-perspective-on-the-deep-image","title":"A Bayesian Perspective on the Deep Image Prior","arxiv_id":"1904.07457","date":"2019-04-16","proceeding":"CVPR 2019 6","authors":["Zezhou Cheng","Matheus Gadelha","Subhransu Maji","Daniel Sheldon"],"abstract":"The deep image prior was recently introduced as a prior for natural images.\nIt represents images as the output of a convolutional network with random\ninputs. For \"inference\", gradient descent is performed to adjust network\nparameters to make the output match observations. This approach yields good\nperformance on a range of image reconstruction tasks. We show that the deep\nimage prior is asymptotically equivalent to a stationary Gaussian process prior\nin the limit as the number of channels in each layer of the network goes to\ninfinity, and derive the corresponding kernel. This informs a Bayesian approach\nto inference. We show that by conducting posterior inference using stochastic\ngradient Langevin we avoid the need for early stopping, which is a drawback of\nthe current approach, and improve results for denoising and impainting tasks.\nWe illustrate these intuitions on a number of 1D and 2D signal reconstruction\ntasks.","url_abs":"http://arxiv.org/abs/1904.07457v1","url_pdf":"http://arxiv.org/pdf/1904.07457v1.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":"a-bayesian-perspective-on-the-deep-image","repo_url":"https://github.com/ZezhouCheng/GP-DIP","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":{"status":"ok"}}],"tasks":[{"task_slug":"bayesian-inference","task_name":"Bayesian Inference"},{"task_slug":"denoising","task_name":"Denoising"},{"task_slug":"image-reconstruction","task_name":"Image Reconstruction"}],"methods":[{"method_slug":"gaussian-process","method_name":"Gaussian Process"}],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1904.07457","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}