{"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/h-detach-modifying-the-lstm-gradient-towards","title":"h-detach: Modifying the LSTM Gradient Towards Better Optimization","arxiv_id":"1810.03023","date":"2018-10-06","proceeding":"ICLR 2019 5","authors":["Devansh Arpit","Bhargav Kanuparthi","Giancarlo Kerg","Nan Rosemary Ke","Ioannis Mitliagkas","Yoshua Bengio"],"abstract":"Recurrent neural networks are known for their notorious exploding and\nvanishing gradient problem (EVGP). This problem becomes more evident in tasks\nwhere the information needed to correctly solve them exist over long time\nscales, because EVGP prevents important gradient components from being\nback-propagated adequately over a large number of steps. We introduce a simple\nstochastic algorithm (\\textit{h}-detach) that is specific to LSTM optimization\nand targeted towards addressing this problem. Specifically, we show that when\nthe LSTM weights are large, the gradient components through the linear path\n(cell state) in the LSTM computational graph get suppressed. Based on the\nhypothesis that these components carry information about long term dependencies\n(which we show empirically), their suppression can prevent LSTMs from capturing\nthem. Our algorithm\\footnote{Our code is available at\nhttps://github.com/bhargav104/h-detach.} prevents gradients flowing through\nthis path from getting suppressed, thus allowing the LSTM to capture such\ndependencies better. We show significant improvements over vanilla LSTM\ngradient based training in terms of convergence speed, robustness to seed and\nlearning rate, and generalization using our modification of LSTM gradient on\nvarious benchmark datasets.","url_abs":"http://arxiv.org/abs/1810.03023v2","url_pdf":"http://arxiv.org/pdf/1810.03023v2.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":"h-detach-modifying-the-lstm-gradient-towards","repo_url":"https://github.com/bhargav104/h-detach","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":1,"framework":"pytorch","reach":{"status":"ok","spdx":"NOASSERTION"}}],"tasks":[],"methods":[{"method_slug":"lstm","method_name":"LSTM"},{"method_slug":"sigmoid-activation","method_name":"Sigmoid Activation"},{"method_slug":"tanh-activation","method_name":"Tanh Activation"}],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"syntology_url":"https://syntology.ai/paper/1810.03023","atlas_url":"https://app.syntology.ai/?focus=1810.03023","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}