Generalization without Systematicity: On the Compositional Skills of Sequence-to-Sequence Recurrent Networks. Lake, B. & Baroni, M. In Proceedings of the 35th International Conference on Machine Learning, pages 2873–2882, July, 2018. PMLR. ![Generalization without Systematicity: On the Compositional Skills of Sequence-to-Sequence Recurrent Networks [link]](https://bibbase.org/img/filetypes/link.svg "link")
Paper abstract bibtex 8 downloads Humans can understand and produce new utterances effortlessly, thanks to their compositional skills. Once a person learns the meaning of a new verb "dax," he or she can immediately understand the meaning of "dax twice" or "sing and dax." In this paper, we introduce the SCAN domain, consisting of a set of simple compositional navigation commands paired with the corresponding action sequences. We then test the zero-shot generalization capabilities of a variety of recurrent neural networks (RNNs) trained on SCAN with sequence-to-sequence methods. We find that RNNs can make successful zero-shot generalizations when the differences between training and test commands are small, so that they can apply "mix-and-match" strategies to solve the task. However, when generalization requires systematic compositional skills (as in the "dax" example above), RNNs fail spectacularly. We conclude with a proof-of-concept experiment in neural machine translation, suggesting that lack of systematicity might be partially responsible for neural networks' notorious training data thirst.
@inproceedings{lakeGeneralizationSystematicityCompositional2018,
title = {Generalization without {{Systematicity}}: {{On}} the {{Compositional Skills}} of {{Sequence-to-Sequence Recurrent Networks}}},
shorttitle = {Generalization without {{Systematicity}}},
booktitle = {Proceedings of the 35th {{International Conference}} on {{Machine Learning}}},
author = {Lake, Brenden and Baroni, Marco},
year = {2018},
month = jul,
pages = {2873--2882},
publisher = {PMLR},
issn = {2640-3498},
url = {https://proceedings.mlr.press/v80/lake18a.html},
urldate = {2024-03-18},
abstract = {Humans can understand and produce new utterances effortlessly, thanks to their compositional skills. Once a person learns the meaning of a new verb "dax," he or she can immediately understand the meaning of "dax twice" or "sing and dax." In this paper, we introduce the SCAN domain, consisting of a set of simple compositional navigation commands paired with the corresponding action sequences. We then test the zero-shot generalization capabilities of a variety of recurrent neural networks (RNNs) trained on SCAN with sequence-to-sequence methods. We find that RNNs can make successful zero-shot generalizations when the differences between training and test commands are small, so that they can apply "mix-and-match" strategies to solve the task. However, when generalization requires systematic compositional skills (as in the "dax" example above), RNNs fail spectacularly. We conclude with a proof-of-concept experiment in neural machine translation, suggesting that lack of systematicity might be partially responsible for neural networks' notorious training data thirst.},
langid = {english},
file = {/Users/shanest/sync/library/Lake_Baroni/2018/Lake and Baroni - 2018 - Generalization without Systematicity On the Compo2.pdf;/Users/shanest/sync/library/Lake_Baroni/2018/Lake and Baroni - 2018 - Generalization without Systematicity On the Compo3.pdf}
}
Downloads: 8
{"_id":"vM9Xa4xQpnkCbzjiZ","bibbaseid":"lake-baroni-generalizationwithoutsystematicityonthecompositionalskillsofsequencetosequencerecurrentnetworks-2018","authorIDs":[],"author_short":["Lake, B.","Baroni, M."],"bibdata":{"bibtype":"inproceedings","type":"inproceedings","title":"Generalization without Systematicity: On the Compositional Skills of Sequence-to-Sequence Recurrent Networks","shorttitle":"Generalization without Systematicity","booktitle":"Proceedings of the 35th International Conference on Machine Learning","author":[{"propositions":[],"lastnames":["Lake"],"firstnames":["Brenden"],"suffixes":[]},{"propositions":[],"lastnames":["Baroni"],"firstnames":["Marco"],"suffixes":[]}],"year":"2018","month":"July","pages":"2873–2882","publisher":"PMLR","issn":"2640-3498","url":"https://proceedings.mlr.press/v80/lake18a.html","urldate":"2024-03-18","abstract":"Humans can understand and produce new utterances effortlessly, thanks to their compositional skills. Once a person learns the meaning of a new verb \"dax,\" he or she can immediately understand the meaning of \"dax twice\" or \"sing and dax.\" In this paper, we introduce the SCAN domain, consisting of a set of simple compositional navigation commands paired with the corresponding action sequences. We then test the zero-shot generalization capabilities of a variety of recurrent neural networks (RNNs) trained on SCAN with sequence-to-sequence methods. We find that RNNs can make successful zero-shot generalizations when the differences between training and test commands are small, so that they can apply \"mix-and-match\" strategies to solve the task. However, when generalization requires systematic compositional skills (as in the \"dax\" example above), RNNs fail spectacularly. We conclude with a proof-of-concept experiment in neural machine translation, suggesting that lack of systematicity might be partially responsible for neural networks' notorious training data thirst.","langid":"english","file":"/Users/shanest/sync/library/Lake_Baroni/2018/Lake and Baroni - 2018 - Generalization without Systematicity On the Compo2.pdf;/Users/shanest/sync/library/Lake_Baroni/2018/Lake and Baroni - 2018 - Generalization without Systematicity On the Compo3.pdf","bibtex":"@inproceedings{lakeGeneralizationSystematicityCompositional2018,\n title = {Generalization without {{Systematicity}}: {{On}} the {{Compositional Skills}} of {{Sequence-to-Sequence Recurrent Networks}}},\n shorttitle = {Generalization without {{Systematicity}}},\n booktitle = {Proceedings of the 35th {{International Conference}} on {{Machine Learning}}},\n author = {Lake, Brenden and Baroni, Marco},\n year = {2018},\n month = jul,\n pages = {2873--2882},\n publisher = {PMLR},\n issn = {2640-3498},\n url = {https://proceedings.mlr.press/v80/lake18a.html},\n urldate = {2024-03-18},\n abstract = {Humans can understand and produce new utterances effortlessly, thanks to their compositional skills. Once a person learns the meaning of a new verb \"dax,\" he or she can immediately understand the meaning of \"dax twice\" or \"sing and dax.\" In this paper, we introduce the SCAN domain, consisting of a set of simple compositional navigation commands paired with the corresponding action sequences. We then test the zero-shot generalization capabilities of a variety of recurrent neural networks (RNNs) trained on SCAN with sequence-to-sequence methods. We find that RNNs can make successful zero-shot generalizations when the differences between training and test commands are small, so that they can apply \"mix-and-match\" strategies to solve the task. However, when generalization requires systematic compositional skills (as in the \"dax\" example above), RNNs fail spectacularly. We conclude with a proof-of-concept experiment in neural machine translation, suggesting that lack of systematicity might be partially responsible for neural networks' notorious training data thirst.},\n langid = {english},\n file = {/Users/shanest/sync/library/Lake_Baroni/2018/Lake and Baroni - 2018 - Generalization without Systematicity On the Compo2.pdf;/Users/shanest/sync/library/Lake_Baroni/2018/Lake and Baroni - 2018 - Generalization without Systematicity On the Compo3.pdf}\n}\n\n","author_short":["Lake, B.","Baroni, M."],"key":"lakeGeneralizationSystematicityCompositional2018","id":"lakeGeneralizationSystematicityCompositional2018","bibbaseid":"lake-baroni-generalizationwithoutsystematicityonthecompositionalskillsofsequencetosequencerecurrentnetworks-2018","role":"author","urls":{"Paper":"https://proceedings.mlr.press/v80/lake18a.html"},"metadata":{"authorlinks":{}},"downloads":8},"bibtype":"inproceedings","biburl":"https://www.shane.st/teaching/575/spr24/575_Compositionality.bib","creationDate":"2020-01-05T04:04:02.892Z","downloads":10,"keywords":[],"search_terms":["generalization","without","systematicity","compositional","skills","sequence","sequence","recurrent","networks","lake","baroni"],"title":"Generalization without Systematicity: On the Compositional Skills of Sequence-to-Sequence Recurrent Networks","year":2018,"dataSources":["okYcdTpf4JJ2zkj7A","TPZs4iPqAgE5a8mjq","znj7izS5PeehdLR3G"]}