Analyzing Multi-Head Self-Attention: Specialized Heads Do the Heavy Lifting, the Rest Can Be Pruned

Analyzing Multi-Head Self-Attention: Specialized Heads Do the Heavy Lifting, the Rest Can Be Pruned. Voita, E., Talbot, D., Moiseev, F., Sennrich, R., & Titov, I. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5797–5808, Stroudsburg, PA, USA, 2019. Association for Computational Linguistics.

Paper doi abstract bibtex 1 download

Multi-head self-attention is a key component of the Transformer, a state-of-the-art architecture for neural machine translation. In this work we evaluate the contribution made by individual attention heads in the encoder to the overall performance of the model and analyze the roles played by them. We find that the most important and confident heads play consistent and often linguistically-interpretable roles. When pruning heads using a method based on stochastic gates and a differentiable relaxation of the L0 penalty, we observe that specialized heads are last to be pruned. Our novel pruning method removes the vast majority of heads without seriously affecting performance. For example, on the English-Russian WMT dataset, pruning 38 out of 48 encoder heads results in a drop of only 0.15 BLEU.

@inproceedings{Voita2019a,
abstract = {Multi-head self-attention is a key component of the Transformer, a state-of-the-art architecture for neural machine translation. In this work we evaluate the contribution made by individual attention heads in the encoder to the overall performance of the model and analyze the roles played by them. We find that the most important and confident heads play consistent and often linguistically-interpretable roles. When pruning heads using a method based on stochastic gates and a differentiable relaxation of the L0 penalty, we observe that specialized heads are last to be pruned. Our novel pruning method removes the vast majority of heads without seriously affecting performance. For example, on the English-Russian WMT dataset, pruning 38 out of 48 encoder heads results in a drop of only 0.15 BLEU.},
address = {Stroudsburg, PA, USA},
archivePrefix = {arXiv},
arxivId = {1905.09418},
author = {Voita, Elena and Talbot, David and Moiseev, Fedor and Sennrich, Rico and Titov, Ivan},
booktitle = {Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics},
doi = {10.18653/v1/P19-1580},
eprint = {1905.09418},
file = {:Users/shanest/Documents/Library/Voita et al/Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics/Voita et al. - 2019 - Analyzing Multi-Head Self-Attention Specialized Heads Do the Heavy Lifting, the Rest Can Be Pruned.pdf:pdf},
keywords = {method: attention,method: pruning},
pages = {5797--5808},
publisher = {Association for Computational Linguistics},
title = {{Analyzing Multi-Head Self-Attention: Specialized Heads Do the Heavy Lifting, the Rest Can Be Pruned}},
url = {https://www.aclweb.org/anthology/P19-1580},
year = {2019}
}

Downloads: 1

{"_id":"qW4Zpj33JuQxeR3bi","bibbaseid":"voita-talbot-moiseev-sennrich-titov-analyzingmultiheadselfattentionspecializedheadsdotheheavyliftingtherestcanbepruned-2019","authorIDs":[],"author_short":["Voita, E.","Talbot, D.","Moiseev, F.","Sennrich, R.","Titov, I."],"bibdata":{"bibtype":"inproceedings","type":"inproceedings","abstract":"Multi-head self-attention is a key component of the Transformer, a state-of-the-art architecture for neural machine translation. In this work we evaluate the contribution made by individual attention heads in the encoder to the overall performance of the model and analyze the roles played by them. We find that the most important and confident heads play consistent and often linguistically-interpretable roles. When pruning heads using a method based on stochastic gates and a differentiable relaxation of the L0 penalty, we observe that specialized heads are last to be pruned. Our novel pruning method removes the vast majority of heads without seriously affecting performance. For example, on the English-Russian WMT dataset, pruning 38 out of 48 encoder heads results in a drop of only 0.15 BLEU.","address":"Stroudsburg, PA, USA","archiveprefix":"arXiv","arxivid":"1905.09418","author":[{"propositions":[],"lastnames":["Voita"],"firstnames":["Elena"],"suffixes":[]},{"propositions":[],"lastnames":["Talbot"],"firstnames":["David"],"suffixes":[]},{"propositions":[],"lastnames":["Moiseev"],"firstnames":["Fedor"],"suffixes":[]},{"propositions":[],"lastnames":["Sennrich"],"firstnames":["Rico"],"suffixes":[]},{"propositions":[],"lastnames":["Titov"],"firstnames":["Ivan"],"suffixes":[]}],"booktitle":"Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics","doi":"10.18653/v1/P19-1580","eprint":"1905.09418","file":":Users/shanest/Documents/Library/Voita et al/Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics/Voita et al. - 2019 - Analyzing Multi-Head Self-Attention Specialized Heads Do the Heavy Lifting, the Rest Can Be Pruned.pdf:pdf","keywords":"method: attention,method: pruning","pages":"5797–5808","publisher":"Association for Computational Linguistics","title":"Analyzing Multi-Head Self-Attention: Specialized Heads Do the Heavy Lifting, the Rest Can Be Pruned","url":"https://www.aclweb.org/anthology/P19-1580","year":"2019","bibtex":"@inproceedings{Voita2019a,\nabstract = {Multi-head self-attention is a key component of the Transformer, a state-of-the-art architecture for neural machine translation. In this work we evaluate the contribution made by individual attention heads in the encoder to the overall performance of the model and analyze the roles played by them. We find that the most important and confident heads play consistent and often linguistically-interpretable roles. When pruning heads using a method based on stochastic gates and a differentiable relaxation of the L0 penalty, we observe that specialized heads are last to be pruned. Our novel pruning method removes the vast majority of heads without seriously affecting performance. For example, on the English-Russian WMT dataset, pruning 38 out of 48 encoder heads results in a drop of only 0.15 BLEU.},\naddress = {Stroudsburg, PA, USA},\narchivePrefix = {arXiv},\narxivId = {1905.09418},\nauthor = {Voita, Elena and Talbot, David and Moiseev, Fedor and Sennrich, Rico and Titov, Ivan},\nbooktitle = {Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics},\ndoi = {10.18653/v1/P19-1580},\neprint = {1905.09418},\nfile = {:Users/shanest/Documents/Library/Voita et al/Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics/Voita et al. - 2019 - Analyzing Multi-Head Self-Attention Specialized Heads Do the Heavy Lifting, the Rest Can Be Pruned.pdf:pdf},\nkeywords = {method: attention,method: pruning},\npages = {5797--5808},\npublisher = {Association for Computational Linguistics},\ntitle = {{Analyzing Multi-Head Self-Attention: Specialized Heads Do the Heavy Lifting, the Rest Can Be Pruned}},\nurl = {https://www.aclweb.org/anthology/P19-1580},\nyear = {2019}\n}\n","author_short":["Voita, E.","Talbot, D.","Moiseev, F.","Sennrich, R.","Titov, I."],"key":"Voita2019a","id":"Voita2019a","bibbaseid":"voita-talbot-moiseev-sennrich-titov-analyzingmultiheadselfattentionspecializedheadsdotheheavyliftingtherestcanbepruned-2019","role":"author","urls":{"Paper":"https://www.aclweb.org/anthology/P19-1580"},"keyword":["method: attention","method: pruning"],"metadata":{"authorlinks":{}},"downloads":1},"bibtype":"inproceedings","biburl":"https://www.shane.st/teaching/575/win20/MachineLearning-interpretability.bib","creationDate":"2020-01-17T05:09:05.721Z","downloads":1,"keywords":["method: attention","method: pruning"],"search_terms":["analyzing","multi","head","self","attention","specialized","heads","heavy","lifting","rest","pruned","voita","talbot","moiseev","sennrich","titov"],"title":"Analyzing Multi-Head Self-Attention: Specialized Heads Do the Heavy Lifting, the Rest Can Be Pruned","year":2019,"dataSources":["okYcdTpf4JJ2zkj7A"]}