@inproceedings{hein-diepold-2022-minimal,
title = "A Minimal Model for Compositional Generalization on g{SCAN}",
author = "Hein, Alice and
Diepold, Klaus",
editor = "Bastings, Jasmijn and
Belinkov, Yonatan and
Elazar, Yanai and
Hupkes, Dieuwke and
Saphra, Naomi and
Wiegreffe, Sarah",
booktitle = "Proceedings of the Fifth BlackboxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP",
month = dec,
year = "2022",
address = "Abu Dhabi, United Arab Emirates (Hybrid)",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2022.blackboxnlp-1.1",
doi = "10.18653/v1/2022.blackboxnlp-1.1",
pages = "1--15",
abstract = "Whether neural networks are capable of compositional generalization has been a topic of much debate. Most previous studies on this subject investigate the generalization capabilities of state-of-the-art deep learning architectures. We here take a more bottom-up approach and design a minimal model that displays generalization on a compositional benchmark, namely, the gSCAN dataset. The model is a hybrid architecture that combines layers trained with gradient descent and a selective attention mechanism optimized with an evolutionary strategy. The architecture has around 60 times fewer trainable parameters than models previously tested on gSCAN, and achieves comparable accuracies on most test splits, even when trained only on a fraction of the dataset. On adverb to verb generalization accuracy, it outperforms previous approaches by 65 to 86{\%}. Through ablation studies, neuron pruning, and error analyses, we show that weight decay and attention mechanisms facilitate compositional generalization by encouraging sparse representations divorced from irrelevant context. We find that the model{'}s sample efficiency can mainly be attributed to its selective attention mechanism.",
}
<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="hein-diepold-2022-minimal">
<titleInfo>
<title>A Minimal Model for Compositional Generalization on gSCAN</title>
</titleInfo>
<name type="personal">
<namePart type="given">Alice</namePart>
<namePart type="family">Hein</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Klaus</namePart>
<namePart type="family">Diepold</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2022-12</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the Fifth BlackboxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP</title>
</titleInfo>
<name type="personal">
<namePart type="given">Jasmijn</namePart>
<namePart type="family">Bastings</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yonatan</namePart>
<namePart type="family">Belinkov</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yanai</namePart>
<namePart type="family">Elazar</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Dieuwke</namePart>
<namePart type="family">Hupkes</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Naomi</namePart>
<namePart type="family">Saphra</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sarah</namePart>
<namePart type="family">Wiegreffe</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Abu Dhabi, United Arab Emirates (Hybrid)</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>Whether neural networks are capable of compositional generalization has been a topic of much debate. Most previous studies on this subject investigate the generalization capabilities of state-of-the-art deep learning architectures. We here take a more bottom-up approach and design a minimal model that displays generalization on a compositional benchmark, namely, the gSCAN dataset. The model is a hybrid architecture that combines layers trained with gradient descent and a selective attention mechanism optimized with an evolutionary strategy. The architecture has around 60 times fewer trainable parameters than models previously tested on gSCAN, and achieves comparable accuracies on most test splits, even when trained only on a fraction of the dataset. On adverb to verb generalization accuracy, it outperforms previous approaches by 65 to 86%. Through ablation studies, neuron pruning, and error analyses, we show that weight decay and attention mechanisms facilitate compositional generalization by encouraging sparse representations divorced from irrelevant context. We find that the model’s sample efficiency can mainly be attributed to its selective attention mechanism.</abstract>
<identifier type="citekey">hein-diepold-2022-minimal</identifier>
<identifier type="doi">10.18653/v1/2022.blackboxnlp-1.1</identifier>
<location>
<url>https://aclanthology.org/2022.blackboxnlp-1.1</url>
</location>
<part>
<date>2022-12</date>
<extent unit="page">
<start>1</start>
<end>15</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T A Minimal Model for Compositional Generalization on gSCAN
%A Hein, Alice
%A Diepold, Klaus
%Y Bastings, Jasmijn
%Y Belinkov, Yonatan
%Y Elazar, Yanai
%Y Hupkes, Dieuwke
%Y Saphra, Naomi
%Y Wiegreffe, Sarah
%S Proceedings of the Fifth BlackboxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP
%D 2022
%8 December
%I Association for Computational Linguistics
%C Abu Dhabi, United Arab Emirates (Hybrid)
%F hein-diepold-2022-minimal
%X Whether neural networks are capable of compositional generalization has been a topic of much debate. Most previous studies on this subject investigate the generalization capabilities of state-of-the-art deep learning architectures. We here take a more bottom-up approach and design a minimal model that displays generalization on a compositional benchmark, namely, the gSCAN dataset. The model is a hybrid architecture that combines layers trained with gradient descent and a selective attention mechanism optimized with an evolutionary strategy. The architecture has around 60 times fewer trainable parameters than models previously tested on gSCAN, and achieves comparable accuracies on most test splits, even when trained only on a fraction of the dataset. On adverb to verb generalization accuracy, it outperforms previous approaches by 65 to 86%. Through ablation studies, neuron pruning, and error analyses, we show that weight decay and attention mechanisms facilitate compositional generalization by encouraging sparse representations divorced from irrelevant context. We find that the model’s sample efficiency can mainly be attributed to its selective attention mechanism.
%R 10.18653/v1/2022.blackboxnlp-1.1
%U https://aclanthology.org/2022.blackboxnlp-1.1
%U https://doi.org/10.18653/v1/2022.blackboxnlp-1.1
%P 1-15
Markdown (Informal)
[A Minimal Model for Compositional Generalization on gSCAN](https://aclanthology.org/2022.blackboxnlp-1.1) (Hein & Diepold, BlackboxNLP 2022)
ACL
- Alice Hein and Klaus Diepold. 2022. A Minimal Model for Compositional Generalization on gSCAN. In Proceedings of the Fifth BlackboxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP, pages 1–15, Abu Dhabi, United Arab Emirates (Hybrid). Association for Computational Linguistics.