{"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/full-capacity-unitary-recurrent-neural","title":"Full-Capacity Unitary Recurrent Neural Networks","arxiv_id":"1611.00035","date":"2016-10-31","proceeding":"NeurIPS 2016 12","authors":["Scott Wisdom","Thomas Powers","John R. Hershey","Jonathan Le Roux","Les Atlas"],"abstract":"Recurrent neural networks are powerful models for processing sequential data,\nbut they are generally plagued by vanishing and exploding gradient problems.\nUnitary recurrent neural networks (uRNNs), which use unitary recurrence\nmatrices, have recently been proposed as a means to avoid these issues.\nHowever, in previous experiments, the recurrence matrices were restricted to be\na product of parameterized unitary matrices, and an open question remains: when\ndoes such a parameterization fail to represent all unitary matrices, and how\ndoes this restricted representational capacity limit what can be learned? To\naddress this question, we propose full-capacity uRNNs that optimize their\nrecurrence matrix over all unitary matrices, leading to significantly improved\nperformance over uRNNs that use a restricted-capacity recurrence matrix. Our\ncontribution consists of two main components. First, we provide a theoretical\nargument to determine if a unitary parameterization has restricted capacity.\nUsing this argument, we show that a recently proposed unitary parameterization\nhas restricted capacity for hidden state dimension greater than 7. Second, we\nshow how a complete, full-capacity unitary recurrence matrix can be optimized\nover the differentiable manifold of unitary matrices. The resulting\nmultiplicative gradient step is very simple and does not require gradient\nclipping or learning rate adaptation. We confirm the utility of our claims by\nempirically evaluating our new full-capacity uRNNs on both synthetic and\nnatural data, achieving superior performance compared to both LSTMs and the\noriginal restricted-capacity uRNNs.","url_abs":"http://arxiv.org/abs/1611.00035v1","url_pdf":"http://arxiv.org/pdf/1611.00035v1.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":"full-capacity-unitary-recurrent-neural","repo_url":"https://github.com/stwisdom/urnn","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":0,"framework":"none","reach":{"status":"unanswered"}},{"paper_slug":"full-capacity-unitary-recurrent-neural","repo_url":"https://github.com/v0lta/Complex-gated-recurrent-neural-networks","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"tf","reach":null}],"tasks":[{"task_slug":"open-question","task_name":"Open-Ended Question Answering"},{"task_slug":"sequential-image-classification","task_name":"Sequential Image Classification"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[{"leaderboard":"/sota/sequential-image-classification-on-sequential","task":"Sequential Image Classification","dataset":"Sequential MNIST","model":"Full-capacity uRNN","rank_in_archive_order":25,"of":30,"metrics":{"Permuted Accuracy":"94.1%","Unpermuted Accuracy":"96.9%"},"uses_additional_data":false}],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1611.00035","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}