<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="puccetti-etal-2022-outlier">
    <titleInfo>
        <title>Outlier Dimensions that Disrupt Transformers are Driven by Frequency</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Giovanni</namePart>
        <namePart type="family">Puccetti</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Anna</namePart>
        <namePart type="family">Rogers</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Aleksandr</namePart>
        <namePart type="family">Drozd</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Felice</namePart>
        <namePart type="family">Dell’Orletta</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>Findings of the Association for Computational Linguistics: EMNLP 2022</title>
        </titleInfo>
        <name type="personal">
            <namePart type="given">Yoav</namePart>
            <namePart type="family">Goldberg</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Zornitsa</namePart>
            <namePart type="family">Kozareva</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Yue</namePart>
            <namePart type="family">Zhang</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</placeTerm>
            </place>
        </originInfo>
        <genre authority="marcgt">conference publication</genre>
    </relatedItem>
    <abstract>While Transformer-based language models are generally very robust to pruning, there is the recently discovered outlier phenomenon: disabling only 48 out of 110M parameters in BERT-base drops its performance by nearly 30% on MNLI. We replicate the original evidence for the outlier phenomenon and we link it to the geometry of the embedding space. We find that in both BERT and RoBERTa the magnitude of hidden state coefficients corresponding to outlier dimensions correlate with the frequencies of encoded tokens in pre-training data, and they also contribute to the “vertical” self-attention pattern enabling the model to focus on the special tokens. This explains the drop in performance from disabling the outliers, and it suggests that to decrease anisotopicity in future models we need pre-training schemas that would better take into account the skewed token distributions.</abstract>
    <identifier type="citekey">puccetti-etal-2022-outlier</identifier>
    <identifier type="doi">10.18653/v1/2022.findings-emnlp.93</identifier>
    <location>
        <url>https://aclanthology.org/2022.findings-emnlp.93</url>
    </location>
    <part>
        <date>2022-12</date>
        <extent unit="page">
            <start>1286</start>
            <end>1304</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T Outlier Dimensions that Disrupt Transformers are Driven by Frequency
%A Puccetti, Giovanni
%A Rogers, Anna
%A Drozd, Aleksandr
%A Dell’Orletta, Felice
%Y Goldberg, Yoav
%Y Kozareva, Zornitsa
%Y Zhang, Yue
%S Findings of the Association for Computational Linguistics: EMNLP 2022
%D 2022
%8 December
%I Association for Computational Linguistics
%C Abu Dhabi, United Arab Emirates
%F puccetti-etal-2022-outlier
%X While Transformer-based language models are generally very robust to pruning, there is the recently discovered outlier phenomenon: disabling only 48 out of 110M parameters in BERT-base drops its performance by nearly 30% on MNLI. We replicate the original evidence for the outlier phenomenon and we link it to the geometry of the embedding space. We find that in both BERT and RoBERTa the magnitude of hidden state coefficients corresponding to outlier dimensions correlate with the frequencies of encoded tokens in pre-training data, and they also contribute to the “vertical” self-attention pattern enabling the model to focus on the special tokens. This explains the drop in performance from disabling the outliers, and it suggests that to decrease anisotopicity in future models we need pre-training schemas that would better take into account the skewed token distributions.
%R 10.18653/v1/2022.findings-emnlp.93
%U https://aclanthology.org/2022.findings-emnlp.93
%U https://doi.org/10.18653/v1/2022.findings-emnlp.93
%P 1286-1304

Markdown (Informal)

[Outlier Dimensions that Disrupt Transformers are Driven by Frequency](https://aclanthology.org/2022.findings-emnlp.93) (Puccetti et al., Findings 2022)

• Outlier Dimensions that Disrupt Transformers are Driven by Frequency (Puccetti et al., Findings 2022)

ACL

• Giovanni Puccetti, Anna Rogers, Aleksandr Drozd, and Felice Dell’Orletta. 2022. Outlier Dimensions that Disrupt Transformers are Driven by Frequency. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 1286–1304, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.