@inproceedings{lu-etal-2022-asynchronous,
title = "Asynchronous Convergence in Multi-Task Learning via Knowledge Distillation from Converged Tasks",
author = "Lu, Weiyi and
Rajagopalan, Sunny and
Nigam, Priyanka and
Singh, Jaspreet and
Sun, Xiaodi and
Xu, Yi and
Zeng, Belinda and
Chilimbi, Trishul",
editor = "Loukina, Anastassia and
Gangadharaiah, Rashmi and
Min, Bonan",
booktitle = "Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies: Industry Track",
month = jul,
year = "2022",
address = "Hybrid: Seattle, Washington + Online",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2022.naacl-industry.18/",
doi = "10.18653/v1/2022.naacl-industry.18",
pages = "149--159",
abstract = "Multi-task learning (MTL) aims to solve multiple tasks jointly by sharing a base representation among them. This can lead to more efficient learning and better generalization, as compared to learning each task individually. However, one issue that often arises in MTL is the convergence speed between tasks varies due to differences in task difficulty, so it can be a challenge to simultaneously achieve the best performance on all tasks with a single model checkpoint. Various techniques have been proposed to address discrepancies in task convergence rate, including weighting the per-task losses and modifying task gradients. In this work, we propose a novel approach that avoids the problem of requiring all tasks to converge at the same rate, but rather allows for {\textquotedblleft}asynchronous{\textquotedblright} convergence among the tasks where each task can converge on its own schedule. As our main contribution, we monitor per-task validation metrics and switch to a knowledge distillation loss once a task has converged instead of continuing to train on the true labels. This prevents the model from overfitting on converged tasks while it learns the remaining tasks. We evaluate the proposed method in two 5-task MTL setups consisting of internal e-commerce datasets. The results show that our method consistently outperforms existing loss weighting and gradient balancing approaches, achieving average improvements of 0.9{\%} and 1.5{\%} over the best performing baseline model in the two setups, respectively."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="lu-etal-2022-asynchronous">
<titleInfo>
<title>Asynchronous Convergence in Multi-Task Learning via Knowledge Distillation from Converged Tasks</title>
</titleInfo>
<name type="personal">
<namePart type="given">Weiyi</namePart>
<namePart type="family">Lu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunny</namePart>
<namePart type="family">Rajagopalan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Priyanka</namePart>
<namePart type="family">Nigam</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jaspreet</namePart>
<namePart type="family">Singh</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiaodi</namePart>
<namePart type="family">Sun</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yi</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Belinda</namePart>
<namePart type="family">Zeng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Trishul</namePart>
<namePart type="family">Chilimbi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2022-07</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies: Industry Track</title>
</titleInfo>
<name type="personal">
<namePart type="given">Anastassia</namePart>
<namePart type="family">Loukina</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Rashmi</namePart>
<namePart type="family">Gangadharaiah</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Bonan</namePart>
<namePart type="family">Min</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Hybrid: Seattle, Washington + Online</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>Multi-task learning (MTL) aims to solve multiple tasks jointly by sharing a base representation among them. This can lead to more efficient learning and better generalization, as compared to learning each task individually. However, one issue that often arises in MTL is the convergence speed between tasks varies due to differences in task difficulty, so it can be a challenge to simultaneously achieve the best performance on all tasks with a single model checkpoint. Various techniques have been proposed to address discrepancies in task convergence rate, including weighting the per-task losses and modifying task gradients. In this work, we propose a novel approach that avoids the problem of requiring all tasks to converge at the same rate, but rather allows for “asynchronous” convergence among the tasks where each task can converge on its own schedule. As our main contribution, we monitor per-task validation metrics and switch to a knowledge distillation loss once a task has converged instead of continuing to train on the true labels. This prevents the model from overfitting on converged tasks while it learns the remaining tasks. We evaluate the proposed method in two 5-task MTL setups consisting of internal e-commerce datasets. The results show that our method consistently outperforms existing loss weighting and gradient balancing approaches, achieving average improvements of 0.9% and 1.5% over the best performing baseline model in the two setups, respectively.</abstract>
<identifier type="citekey">lu-etal-2022-asynchronous</identifier>
<identifier type="doi">10.18653/v1/2022.naacl-industry.18</identifier>
<location>
<url>https://aclanthology.org/2022.naacl-industry.18/</url>
</location>
<part>
<date>2022-07</date>
<extent unit="page">
<start>149</start>
<end>159</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Asynchronous Convergence in Multi-Task Learning via Knowledge Distillation from Converged Tasks
%A Lu, Weiyi
%A Rajagopalan, Sunny
%A Nigam, Priyanka
%A Singh, Jaspreet
%A Sun, Xiaodi
%A Xu, Yi
%A Zeng, Belinda
%A Chilimbi, Trishul
%Y Loukina, Anastassia
%Y Gangadharaiah, Rashmi
%Y Min, Bonan
%S Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies: Industry Track
%D 2022
%8 July
%I Association for Computational Linguistics
%C Hybrid: Seattle, Washington + Online
%F lu-etal-2022-asynchronous
%X Multi-task learning (MTL) aims to solve multiple tasks jointly by sharing a base representation among them. This can lead to more efficient learning and better generalization, as compared to learning each task individually. However, one issue that often arises in MTL is the convergence speed between tasks varies due to differences in task difficulty, so it can be a challenge to simultaneously achieve the best performance on all tasks with a single model checkpoint. Various techniques have been proposed to address discrepancies in task convergence rate, including weighting the per-task losses and modifying task gradients. In this work, we propose a novel approach that avoids the problem of requiring all tasks to converge at the same rate, but rather allows for “asynchronous” convergence among the tasks where each task can converge on its own schedule. As our main contribution, we monitor per-task validation metrics and switch to a knowledge distillation loss once a task has converged instead of continuing to train on the true labels. This prevents the model from overfitting on converged tasks while it learns the remaining tasks. We evaluate the proposed method in two 5-task MTL setups consisting of internal e-commerce datasets. The results show that our method consistently outperforms existing loss weighting and gradient balancing approaches, achieving average improvements of 0.9% and 1.5% over the best performing baseline model in the two setups, respectively.
%R 10.18653/v1/2022.naacl-industry.18
%U https://aclanthology.org/2022.naacl-industry.18/
%U https://doi.org/10.18653/v1/2022.naacl-industry.18
%P 149-159
Markdown (Informal)
[Asynchronous Convergence in Multi-Task Learning via Knowledge Distillation from Converged Tasks](https://aclanthology.org/2022.naacl-industry.18/) (Lu et al., NAACL 2022)
ACL
- Weiyi Lu, Sunny Rajagopalan, Priyanka Nigam, Jaspreet Singh, Xiaodi Sun, Yi Xu, Belinda Zeng, and Trishul Chilimbi. 2022. Asynchronous Convergence in Multi-Task Learning via Knowledge Distillation from Converged Tasks. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies: Industry Track, pages 149–159, Hybrid: Seattle, Washington + Online. Association for Computational Linguistics.