{"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/von-mises-fisher-loss-for-training-sequence","title":"Von Mises-Fisher Loss for Training Sequence to Sequence Models with Continuous Outputs","arxiv_id":"1812.04616","date":"2018-12-10","proceeding":"ICLR 2019 5","authors":["Sachin Kumar","Yulia Tsvetkov"],"abstract":"The Softmax function is used in the final layer of nearly all existing\nsequence-to-sequence models for language generation. However, it is usually the\nslowest layer to compute which limits the vocabulary size to a subset of most\nfrequent types; and it has a large memory footprint. We propose a general\ntechnique for replacing the softmax layer with a continuous embedding layer.\nOur primary innovations are a novel probabilistic loss, and a training and\ninference procedure in which we generate a probability distribution over\npre-trained word embeddings, instead of a multinomial distribution over the\nvocabulary obtained via softmax. We evaluate this new class of\nsequence-to-sequence models with continuous outputs on the task of neural\nmachine translation. We show that our models obtain upto 2.5x speed-up in\ntraining time while performing on par with the state-of-the-art models in terms\nof translation quality. These models are capable of handling very large\nvocabularies without compromising on translation quality. They also produce\nmore meaningful errors than in the softmax-based models, as these errors\ntypically lie in a subspace of the vector space of the reference translations.","url_abs":"http://arxiv.org/abs/1812.04616v3","url_pdf":"http://arxiv.org/pdf/1812.04616v3.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":"von-mises-fisher-loss-for-training-sequence","repo_url":"https://github.com/Sachin19/seq2seq-con","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":1,"framework":"pytorch","reach":null}],"tasks":[{"task_slug":"machine-translation","task_name":"Machine Translation"},{"task_slug":"text-generation","task_name":"Text Generation"},{"task_slug":"translation","task_name":"Translation"},{"task_slug":"word-embeddings","task_name":"Word Embeddings"}],"methods":[{"method_slug":"softmax","method_name":"Softmax"}],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"syntology_url":null,"atlas_url":"https://app.syntology.ai/?focus=1812.04616","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}