{"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/blackout-speeding-up-recurrent-neural-network","title":"BlackOut: Speeding up Recurrent Neural Network Language Models With Very Large Vocabularies","arxiv_id":"1511.06909","date":"2015-11-21","proceeding":null,"authors":["Shihao Ji","S. V. N. Vishwanathan","Nadathur Satish","Michael J. Anderson","Pradeep Dubey"],"abstract":"We propose BlackOut, an approximation algorithm to efficiently train massive\nrecurrent neural network language models (RNNLMs) with million word\nvocabularies. BlackOut is motivated by using a discriminative loss, and we\ndescribe a new sampling strategy which significantly reduces computation while\nimproving stability, sample efficiency, and rate of convergence. One way to\nunderstand BlackOut is to view it as an extension of the DropOut strategy to\nthe output layer, wherein we use a discriminative training loss and a weighted\nsampling scheme. We also establish close connections between BlackOut,\nimportance sampling, and noise contrastive estimation (NCE). Our experiments,\non the recently released one billion word language modeling benchmark,\ndemonstrate scalability and accuracy of BlackOut; we outperform the\nstate-of-the art, and achieve the lowest perplexity scores on this dataset.\nMoreover, unlike other established methods which typically require GPUs or CPU\nclusters, we show that a carefully implemented version of BlackOut requires\nonly 1-10 days on a single machine to train a RNNLM with a million word\nvocabulary and billions of parameters on one billion words. Although we\ndescribe BlackOut in the context of RNNLM training, it can be used to any\nnetworks with large softmax output layers.","url_abs":"http://arxiv.org/abs/1511.06909v7","url_pdf":"http://arxiv.org/pdf/1511.06909v7.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":"blackout-speeding-up-recurrent-neural-network","repo_url":"https://github.com/IntelLabs/rnnlm","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":1,"framework":"none","reach":{"status":"gone","observed_at":"2026-09-18","how":"tree_404+repo_404"}}],"tasks":[{"task_slug":null,"task_name":"CPU"},{"task_slug":"language-modeling","task_name":"Language Modeling"},{"task_slug":"language-modelling","task_name":"Language Modelling"}],"methods":[{"method_slug":"dropout","method_name":"Dropout"},{"method_slug":"softmax","method_name":"Softmax"}],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1511.06909","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}