{"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/generalization-without-systematicity-on-the","title":"Generalization without systematicity: On the compositional skills of sequence-to-sequence recurrent networks","arxiv_id":"1711.00350","date":"2017-10-31","proceeding":"ICML 2018 7","authors":["Brenden M. Lake","Marco Baroni"],"abstract":"Humans can understand and produce new utterances effortlessly, thanks to\ntheir compositional skills. Once a person learns the meaning of a new verb\n\"dax,\" he or she can immediately understand the meaning of \"dax twice\" or \"sing\nand dax.\" In this paper, we introduce the SCAN domain, consisting of a set of\nsimple compositional navigation commands paired with the corresponding action\nsequences. We then test the zero-shot generalization capabilities of a variety\nof recurrent neural networks (RNNs) trained on SCAN with sequence-to-sequence\nmethods. We find that RNNs can make successful zero-shot generalizations when\nthe differences between training and test commands are small, so that they can\napply \"mix-and-match\" strategies to solve the task. However, when\ngeneralization requires systematic compositional skills (as in the \"dax\"\nexample above), RNNs fail spectacularly. We conclude with a proof-of-concept\nexperiment in neural machine translation, suggesting that lack of systematicity\nmight be partially responsible for neural networks' notorious training data\nthirst.","url_abs":"http://arxiv.org/abs/1711.00350v3","url_pdf":"http://arxiv.org/pdf/1711.00350v3.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":"generalization-without-systematicity-on-the","repo_url":"https://github.com/brendenlake/SCAN","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":1,"framework":"none","reach":{"status":"ok","spdx":"NOASSERTION"}},{"paper_slug":"generalization-without-systematicity-on-the","repo_url":"https://github.com/JanAthmer/Compositional-generalization-capabillity-of-Transformer","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":null},{"paper_slug":"generalization-without-systematicity-on-the","repo_url":"https://github.com/aman313/SCAN","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":{"status":"ok","spdx":"NOASSERTION"}},{"paper_slug":"generalization-without-systematicity-on-the","repo_url":"https://github.com/arkilpatel/compositional-generalization-seq2seq","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":{"status":"ok","spdx":"MIT"}},{"paper_slug":"generalization-without-systematicity-on-the","repo_url":"https://github.com/i-machine-think/machine-tasks","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"none","reach":{"status":"ok"}},{"paper_slug":"generalization-without-systematicity-on-the","repo_url":"https://github.com/maxwells-daemons/compositional-learning-experiments","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":{"status":"ok","spdx":"Apache-2.0"}},{"paper_slug":"generalization-without-systematicity-on-the","repo_url":"https://github.com/yoonkim/neural-qcfg","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":null}],"tasks":[{"task_slug":"machine-translation","task_name":"Machine Translation"},{"task_slug":"translation","task_name":"Translation"},{"task_slug":"zero-shot-generalization","task_name":"Zero-shot Generalization"}],"methods":[],"datasets_introduced":[{"slug":"scan","name":"SCAN","full_name":"Simplified versions of the CommAI Navigation tasks"}],"methods_introduced":[],"results":[],"syntology":{"syntology_url":"https://syntology.ai/paper/1711.00350","atlas_url":"https://app.syntology.ai/?focus=1711.00350","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}