{"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/capacity-and-trainability-in-recurrent-neural","title":"Capacity and Trainability in Recurrent Neural Networks","arxiv_id":"1611.09913","date":"2016-11-29","proceeding":null,"authors":["Jasmine Collins","Jascha Sohl-Dickstein","David Sussillo"],"abstract":"Two potential bottlenecks on the expressiveness of recurrent neural networks\n(RNNs) are their ability to store information about the task in their\nparameters, and to store information about the input history in their units. We\nshow experimentally that all common RNN architectures achieve nearly the same\nper-task and per-unit capacity bounds with careful training, for a variety of\ntasks and stacking depths. They can store an amount of task information which\nis linear in the number of parameters, and is approximately 5 bits per\nparameter. They can additionally store approximately one real number from their\ninput history per hidden unit. We further find that for several tasks it is the\nper-task parameter capacity bound that determines performance. These results\nsuggest that many previous results comparing RNN architectures are driven\nprimarily by differences in training effectiveness, rather than differences in\ncapacity. Supporting this observation, we compare training difficulty for\nseveral architectures, and show that vanilla RNNs are far more difficult to\ntrain, yet have slightly higher capacity. Finally, we propose two novel RNN\narchitectures, one of which is easier to train than the LSTM or GRU for deeply\nstacked architectures.","url_abs":"http://arxiv.org/abs/1611.09913v3","url_pdf":"http://arxiv.org/pdf/1611.09913v3.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":"capacity-and-trainability-in-recurrent-neural","repo_url":"https://github.com/trevor-richardson/rnn_zoo","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":null}],"tasks":[],"methods":[{"method_slug":"gru","method_name":"GRU"},{"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":{"atlas_url":"https://app.syntology.ai/?focus=1611.09913","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}