{"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/difference-target-propagation","title":"Difference Target Propagation","arxiv_id":"1412.7525","date":"2014-12-23","proceeding":null,"authors":["Dong-Hyun Lee","Saizheng Zhang","Asja Fischer","Yoshua Bengio"],"abstract":"Back-propagation has been the workhorse of recent successes of deep learning\nbut it relies on infinitesimal effects (partial derivatives) in order to\nperform credit assignment. This could become a serious issue as one considers\ndeeper and more non-linear functions, e.g., consider the extreme case of\nnonlinearity where the relation between parameters and cost is actually\ndiscrete. Inspired by the biological implausibility of back-propagation, a few\napproaches have been proposed in the past that could play a similar credit\nassignment role. In this spirit, we explore a novel approach to credit\nassignment in deep networks that we call target propagation. The main idea is\nto compute targets rather than gradients, at each layer. Like gradients, they\nare propagated backwards. In a way that is related but different from\npreviously proposed proxies for back-propagation which rely on a backwards\nnetwork with symmetric weights, target propagation relies on auto-encoders at\neach layer. Unlike back-propagation, it can be applied even when units exchange\nstochastic bits rather than real numbers. We show that a linear correction for\nthe imperfectness of the auto-encoders, called difference target propagation,\nis very effective to make target propagation actually work, leading to results\ncomparable to back-propagation for deep networks with discrete and continuous\nunits and denoising auto-encoders and achieving state of the art for stochastic\nnetworks.","url_abs":"http://arxiv.org/abs/1412.7525v5","url_pdf":"http://arxiv.org/pdf/1412.7525v5.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":"difference-target-propagation","repo_url":"https://github.com/donghyunlee/dtp","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":0,"framework":"none","reach":{"status":"ok"}}],"tasks":[],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1412.7525","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}