{"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/fastgrnn-a-fast-accurate-stable-and-tiny","title":"FastGRNN: A Fast, Accurate, Stable and Tiny Kilobyte Sized Gated Recurrent Neural Network","arxiv_id":"1901.02358","date":"2019-01-08","proceeding":"NeurIPS 2018 12","authors":["Aditya Kusupati","Manish Singh","Kush Bhatia","Ashish Kumar","Prateek Jain","Manik Varma"],"abstract":"This paper develops the FastRNN and FastGRNN algorithms to address the twin\nRNN limitations of inaccurate training and inefficient prediction. Previous\napproaches have improved accuracy at the expense of prediction costs making\nthem infeasible for resource-constrained and real-time applications. Unitary\nRNNs have increased accuracy somewhat by restricting the range of the state\ntransition matrix's singular values but have also increased the model size as\nthey require a larger number of hidden units to make up for the loss in\nexpressive power. Gated RNNs have obtained state-of-the-art accuracies by\nadding extra parameters thereby resulting in even larger models. FastRNN\naddresses these limitations by adding a residual connection that does not\nconstrain the range of the singular values explicitly and has only two extra\nscalar parameters. FastGRNN then extends the residual connection to a gate by\nreusing the RNN matrices to match state-of-the-art gated RNN accuracies but\nwith a 2-4x smaller model. Enforcing FastGRNN's matrices to be low-rank, sparse\nand quantized resulted in accurate models that could be up to 35x smaller than\nleading gated and unitary RNNs. This allowed FastGRNN to accurately recognize\nthe \"Hey Cortana\" wakeword with a 1 KB model and to be deployed on severely\nresource-constrained IoT microcontrollers too tiny to store other RNN models.\nFastGRNN's code is available at https://github.com/Microsoft/EdgeML/.","url_abs":"http://arxiv.org/abs/1901.02358v1","url_pdf":"http://arxiv.org/pdf/1901.02358v1.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":"fastgrnn-a-fast-accurate-stable-and-tiny","repo_url":"https://github.com/Microsoft/EdgeML","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":0,"framework":"tf","reach":null}],"tasks":[{"task_slug":"action-classification","task_name":"Action Classification"},{"task_slug":"language-modelling","task_name":"Language Modelling"},{"task_slug":"speech-recognition","task_name":"Speech Recognition"},{"task_slug":"time-series","task_name":"Time Series Analysis"},{"task_slug":"time-series-classification","task_name":"Time Series Classification"}],"methods":[{"method_slug":"residual-connection","method_name":"Residual Connection"}],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"syntology_url":null,"atlas_url":"https://app.syntology.ai/?focus=1901.02358","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}