@inproceedings{fang-etal-2023-transformers,
title = "Transformers with Learnable Activation Functions",
author = "Fang, Haishuo and
Lee, Ji-Ung and
Moosavi, Nafise Sadat and
Gurevych, Iryna",
editor = "Vlachos, Andreas and
Augenstein, Isabelle",
booktitle = "Findings of the Association for Computational Linguistics: EACL 2023",
month = may,
year = "2023",
address = "Dubrovnik, Croatia",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2023.findings-eacl.181",
doi = "10.18653/v1/2023.findings-eacl.181",
pages = "2382--2398",
abstract = "Activation functions can have a significant impact on reducing the topological complexity of input data and therefore, improving a model{'}s performance. However, the choice of activation functions is seldom discussed or explored in Transformer-based language models. As a common practice, commonly used activation functions like Gaussian Error Linear Unit (GELU) are chosen beforehand and then remain fixed from pre-training to fine-tuning. In this paper, we investigate the impact of activation functions on Transformer-based models by utilizing rational activation functions (RAFs). In contrast to fixed activation functions (FAF), RAFs are capable of learning the optimal activation functions from data. Our experiments show that the RAF-based Transformer model (RAFT) achieves a better performance than its FAF-based counterpart (). For instance, we find that RAFT outperforms on the GLUE benchmark by 5.71 points when using only 100 training examples and by 2.05 points on SQuAD with all available data. Analyzing the shapes of the learned RAFs further unveils that they vary across different layers and different tasks; opening a promising way to better analyze and understand large, pre-trained language models.",
}
<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="fang-etal-2023-transformers">
<titleInfo>
<title>Transformers with Learnable Activation Functions</title>
</titleInfo>
<name type="personal">
<namePart type="given">Haishuo</namePart>
<namePart type="family">Fang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ji-Ung</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Nafise</namePart>
<namePart type="given">Sadat</namePart>
<namePart type="family">Moosavi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Iryna</namePart>
<namePart type="family">Gurevych</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2023-05</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EACL 2023</title>
</titleInfo>
<name type="personal">
<namePart type="given">Andreas</namePart>
<namePart type="family">Vlachos</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Isabelle</namePart>
<namePart type="family">Augenstein</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Dubrovnik, Croatia</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>Activation functions can have a significant impact on reducing the topological complexity of input data and therefore, improving a model’s performance. However, the choice of activation functions is seldom discussed or explored in Transformer-based language models. As a common practice, commonly used activation functions like Gaussian Error Linear Unit (GELU) are chosen beforehand and then remain fixed from pre-training to fine-tuning. In this paper, we investigate the impact of activation functions on Transformer-based models by utilizing rational activation functions (RAFs). In contrast to fixed activation functions (FAF), RAFs are capable of learning the optimal activation functions from data. Our experiments show that the RAF-based Transformer model (RAFT) achieves a better performance than its FAF-based counterpart (). For instance, we find that RAFT outperforms on the GLUE benchmark by 5.71 points when using only 100 training examples and by 2.05 points on SQuAD with all available data. Analyzing the shapes of the learned RAFs further unveils that they vary across different layers and different tasks; opening a promising way to better analyze and understand large, pre-trained language models.</abstract>
<identifier type="citekey">fang-etal-2023-transformers</identifier>
<identifier type="doi">10.18653/v1/2023.findings-eacl.181</identifier>
<location>
<url>https://aclanthology.org/2023.findings-eacl.181</url>
</location>
<part>
<date>2023-05</date>
<extent unit="page">
<start>2382</start>
<end>2398</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Transformers with Learnable Activation Functions
%A Fang, Haishuo
%A Lee, Ji-Ung
%A Moosavi, Nafise Sadat
%A Gurevych, Iryna
%Y Vlachos, Andreas
%Y Augenstein, Isabelle
%S Findings of the Association for Computational Linguistics: EACL 2023
%D 2023
%8 May
%I Association for Computational Linguistics
%C Dubrovnik, Croatia
%F fang-etal-2023-transformers
%X Activation functions can have a significant impact on reducing the topological complexity of input data and therefore, improving a model’s performance. However, the choice of activation functions is seldom discussed or explored in Transformer-based language models. As a common practice, commonly used activation functions like Gaussian Error Linear Unit (GELU) are chosen beforehand and then remain fixed from pre-training to fine-tuning. In this paper, we investigate the impact of activation functions on Transformer-based models by utilizing rational activation functions (RAFs). In contrast to fixed activation functions (FAF), RAFs are capable of learning the optimal activation functions from data. Our experiments show that the RAF-based Transformer model (RAFT) achieves a better performance than its FAF-based counterpart (). For instance, we find that RAFT outperforms on the GLUE benchmark by 5.71 points when using only 100 training examples and by 2.05 points on SQuAD with all available data. Analyzing the shapes of the learned RAFs further unveils that they vary across different layers and different tasks; opening a promising way to better analyze and understand large, pre-trained language models.
%R 10.18653/v1/2023.findings-eacl.181
%U https://aclanthology.org/2023.findings-eacl.181
%U https://doi.org/10.18653/v1/2023.findings-eacl.181
%P 2382-2398
Markdown (Informal)
[Transformers with Learnable Activation Functions](https://aclanthology.org/2023.findings-eacl.181) (Fang et al., Findings 2023)
ACL
- Haishuo Fang, Ji-Ung Lee, Nafise Sadat Moosavi, and Iryna Gurevych. 2023. Transformers with Learnable Activation Functions. In Findings of the Association for Computational Linguistics: EACL 2023, pages 2382–2398, Dubrovnik, Croatia. Association for Computational Linguistics.