{"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/recurrent-neural-networks-with-limited-1","title":"Recurrent Neural Networks With Limited Numerical Precision","arxiv_id":"1608.06902","date":"2016-08-24","proceeding":null,"authors":["Joachim Ott","Zhouhan Lin","Ying Zhang","Shih-Chii Liu","Yoshua Bengio"],"abstract":"Recurrent Neural Networks (RNNs) produce state-of-art performance on many\nmachine learning tasks but their demand on resources in terms of memory and\ncomputational power are often high. Therefore, there is a great interest in\noptimizing the computations performed with these models especially when\nconsidering development of specialized low-power hardware for deep networks.\nOne way of reducing the computational needs is to limit the numerical precision\nof the network weights and biases. This has led to different proposed rounding\nmethods which have been applied so far to only Convolutional Neural Networks\nand Fully-Connected Networks. This paper addresses the question of how to best\nreduce weight precision during training in the case of RNNs. We present results\nfrom the use of different stochastic and deterministic reduced precision\ntraining methods applied to three major RNN types which are then tested on\nseveral datasets. The results show that the weight binarization methods do not\nwork with the RNNs. However, the stochastic and deterministic ternarization,\nand pow2-ternarization methods gave rise to low-precision RNNs that produce\nsimilar and even higher accuracy on certain datasets therefore providing a path\ntowards training more efficient implementations of RNNs in specialized\nhardware.","url_abs":"http://arxiv.org/abs/1608.06902v2","url_pdf":"http://arxiv.org/pdf/1608.06902v2.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":"recurrent-neural-networks-with-limited-1","repo_url":"https://github.com/ottj/QuantizedRNN","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":0,"framework":"none","reach":null}],"tasks":[{"task_slug":"binarization","task_name":"Binarization"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"syntology_url":"https://syntology.ai/paper/1608.06902","atlas_url":"https://app.syntology.ai/?focus=1608.06902","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}