@inproceedings{behnke-heafield-2021-pruning,
title = "Pruning Neural Machine Translation for Speed Using Group Lasso",
author = "Behnke, Maximiliana and
Heafield, Kenneth",
editor = "Barrault, Loic and
Bojar, Ondrej and
Bougares, Fethi and
Chatterjee, Rajen and
Costa-jussa, Marta R. and
Federmann, Christian and
Fishel, Mark and
Fraser, Alexander and
Freitag, Markus and
Graham, Yvette and
Grundkiewicz, Roman and
Guzman, Paco and
Haddow, Barry and
Huck, Matthias and
Yepes, Antonio Jimeno and
Koehn, Philipp and
Kocmi, Tom and
Martins, Andre and
Morishita, Makoto and
Monz, Christof",
booktitle = "Proceedings of the Sixth Conference on Machine Translation",
month = nov,
year = "2021",
address = "Online",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2021.wmt-1.116",
pages = "1074--1086",
abstract = "Unlike most work on pruning neural networks, we make inference faster. Group lasso regularisation enables pruning entire rows, columns or blocks of parameters that result in a smaller dense network. Because the network is still dense, efficient matrix multiply routines are still used and only minimal software changes are required to support variable layer sizes. Moreover, pruning is applied during training so there is no separate pruning step. Experiments on top of English-{\textgreater}German models, which already have state-of-the-art speed and size, show that two-thirds of feedforward connections can be removed with 0.2 BLEU loss. With 6 decoder layers, the pruned model is 34{\%} faster; with 2 tied decoder layers, the pruned model is 14{\%} faster. Pruning entire heads and feedforward connections in a 12{--}1 encoder-decoder architecture gains an additional 51{\%} speed-up. These push the Pareto frontier with respect to the trade-off between time and quality compared to strong baselines. In the WMT 2021 Efficiency Task, our pruned and quantised models are 1.9{--}2.7x faster at the cost 0.9{--}1.7 BLEU in comparison to the unoptimised baselines. Across language pairs, we see similar sparsity patterns: an ascending or U-shaped distribution in encoder feedforward and attention layers and an ascending distribution in the decoder.",
}
<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="behnke-heafield-2021-pruning">
<titleInfo>
<title>Pruning Neural Machine Translation for Speed Using Group Lasso</title>
</titleInfo>
<name type="personal">
<namePart type="given">Maximiliana</namePart>
<namePart type="family">Behnke</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kenneth</namePart>
<namePart type="family">Heafield</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2021-11</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the Sixth Conference on Machine Translation</title>
</titleInfo>
<name type="personal">
<namePart type="given">Loic</namePart>
<namePart type="family">Barrault</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ondrej</namePart>
<namePart type="family">Bojar</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Fethi</namePart>
<namePart type="family">Bougares</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Rajen</namePart>
<namePart type="family">Chatterjee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Marta</namePart>
<namePart type="given">R</namePart>
<namePart type="family">Costa-jussa</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Christian</namePart>
<namePart type="family">Federmann</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mark</namePart>
<namePart type="family">Fishel</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alexander</namePart>
<namePart type="family">Fraser</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Markus</namePart>
<namePart type="family">Freitag</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yvette</namePart>
<namePart type="family">Graham</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Roman</namePart>
<namePart type="family">Grundkiewicz</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Paco</namePart>
<namePart type="family">Guzman</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Barry</namePart>
<namePart type="family">Haddow</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Matthias</namePart>
<namePart type="family">Huck</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Antonio</namePart>
<namePart type="given">Jimeno</namePart>
<namePart type="family">Yepes</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Philipp</namePart>
<namePart type="family">Koehn</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tom</namePart>
<namePart type="family">Kocmi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Andre</namePart>
<namePart type="family">Martins</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Makoto</namePart>
<namePart type="family">Morishita</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Christof</namePart>
<namePart type="family">Monz</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Online</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>Unlike most work on pruning neural networks, we make inference faster. Group lasso regularisation enables pruning entire rows, columns or blocks of parameters that result in a smaller dense network. Because the network is still dense, efficient matrix multiply routines are still used and only minimal software changes are required to support variable layer sizes. Moreover, pruning is applied during training so there is no separate pruning step. Experiments on top of English-\textgreaterGerman models, which already have state-of-the-art speed and size, show that two-thirds of feedforward connections can be removed with 0.2 BLEU loss. With 6 decoder layers, the pruned model is 34% faster; with 2 tied decoder layers, the pruned model is 14% faster. Pruning entire heads and feedforward connections in a 12–1 encoder-decoder architecture gains an additional 51% speed-up. These push the Pareto frontier with respect to the trade-off between time and quality compared to strong baselines. In the WMT 2021 Efficiency Task, our pruned and quantised models are 1.9–2.7x faster at the cost 0.9–1.7 BLEU in comparison to the unoptimised baselines. Across language pairs, we see similar sparsity patterns: an ascending or U-shaped distribution in encoder feedforward and attention layers and an ascending distribution in the decoder.</abstract>
<identifier type="citekey">behnke-heafield-2021-pruning</identifier>
<location>
<url>https://aclanthology.org/2021.wmt-1.116</url>
</location>
<part>
<date>2021-11</date>
<extent unit="page">
<start>1074</start>
<end>1086</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Pruning Neural Machine Translation for Speed Using Group Lasso
%A Behnke, Maximiliana
%A Heafield, Kenneth
%Y Barrault, Loic
%Y Bojar, Ondrej
%Y Bougares, Fethi
%Y Chatterjee, Rajen
%Y Costa-jussa, Marta R.
%Y Federmann, Christian
%Y Fishel, Mark
%Y Fraser, Alexander
%Y Freitag, Markus
%Y Graham, Yvette
%Y Grundkiewicz, Roman
%Y Guzman, Paco
%Y Haddow, Barry
%Y Huck, Matthias
%Y Yepes, Antonio Jimeno
%Y Koehn, Philipp
%Y Kocmi, Tom
%Y Martins, Andre
%Y Morishita, Makoto
%Y Monz, Christof
%S Proceedings of the Sixth Conference on Machine Translation
%D 2021
%8 November
%I Association for Computational Linguistics
%C Online
%F behnke-heafield-2021-pruning
%X Unlike most work on pruning neural networks, we make inference faster. Group lasso regularisation enables pruning entire rows, columns or blocks of parameters that result in a smaller dense network. Because the network is still dense, efficient matrix multiply routines are still used and only minimal software changes are required to support variable layer sizes. Moreover, pruning is applied during training so there is no separate pruning step. Experiments on top of English-\textgreaterGerman models, which already have state-of-the-art speed and size, show that two-thirds of feedforward connections can be removed with 0.2 BLEU loss. With 6 decoder layers, the pruned model is 34% faster; with 2 tied decoder layers, the pruned model is 14% faster. Pruning entire heads and feedforward connections in a 12–1 encoder-decoder architecture gains an additional 51% speed-up. These push the Pareto frontier with respect to the trade-off between time and quality compared to strong baselines. In the WMT 2021 Efficiency Task, our pruned and quantised models are 1.9–2.7x faster at the cost 0.9–1.7 BLEU in comparison to the unoptimised baselines. Across language pairs, we see similar sparsity patterns: an ascending or U-shaped distribution in encoder feedforward and attention layers and an ascending distribution in the decoder.
%U https://aclanthology.org/2021.wmt-1.116
%P 1074-1086
Markdown (Informal)
[Pruning Neural Machine Translation for Speed Using Group Lasso](https://aclanthology.org/2021.wmt-1.116) (Behnke & Heafield, WMT 2021)
ACL