@inproceedings{rudman-etal-2023-outlier,
title = "Outlier Dimensions Encode Task Specific Knowledge",
author = "Rudman, William and
Chen, Catherine and
Eickhoff, Carsten",
editor = "Bouamor, Houda and
Pino, Juan and
Bali, Kalika",
booktitle = "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing",
month = dec,
year = "2023",
address = "Singapore",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2023.emnlp-main.901",
doi = "10.18653/v1/2023.emnlp-main.901",
pages = "14596--14605",
abstract = "Representations from large language models (LLMs) are known to be dominated by a small subset of dimensions with exceedingly high variance. Previous works have argued that although ablating these outlier dimensions in LLM representations hurts downstream performance, outlier dimensions are detrimental to the representational quality of embeddings. In this study, we investigate how fine-tuning impacts outlier dimensions and show that 1) outlier dimensions that occur in pre-training persist in fine-tuned models and 2) a single outlier dimension can complete downstream tasks with a minimal error rate. Our results suggest that outlier dimensions can encode crucial task-specific knowledge and that the value of a representation in a single outlier dimension drives downstream model decisions.",
}
<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="rudman-etal-2023-outlier">
<titleInfo>
<title>Outlier Dimensions Encode Task Specific Knowledge</title>
</titleInfo>
<name type="personal">
<namePart type="given">William</namePart>
<namePart type="family">Rudman</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Catherine</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Carsten</namePart>
<namePart type="family">Eickhoff</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2023-12</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing</title>
</titleInfo>
<name type="personal">
<namePart type="given">Houda</namePart>
<namePart type="family">Bouamor</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Juan</namePart>
<namePart type="family">Pino</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kalika</namePart>
<namePart type="family">Bali</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Singapore</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>Representations from large language models (LLMs) are known to be dominated by a small subset of dimensions with exceedingly high variance. Previous works have argued that although ablating these outlier dimensions in LLM representations hurts downstream performance, outlier dimensions are detrimental to the representational quality of embeddings. In this study, we investigate how fine-tuning impacts outlier dimensions and show that 1) outlier dimensions that occur in pre-training persist in fine-tuned models and 2) a single outlier dimension can complete downstream tasks with a minimal error rate. Our results suggest that outlier dimensions can encode crucial task-specific knowledge and that the value of a representation in a single outlier dimension drives downstream model decisions.</abstract>
<identifier type="citekey">rudman-etal-2023-outlier</identifier>
<identifier type="doi">10.18653/v1/2023.emnlp-main.901</identifier>
<location>
<url>https://aclanthology.org/2023.emnlp-main.901</url>
</location>
<part>
<date>2023-12</date>
<extent unit="page">
<start>14596</start>
<end>14605</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Outlier Dimensions Encode Task Specific Knowledge
%A Rudman, William
%A Chen, Catherine
%A Eickhoff, Carsten
%Y Bouamor, Houda
%Y Pino, Juan
%Y Bali, Kalika
%S Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing
%D 2023
%8 December
%I Association for Computational Linguistics
%C Singapore
%F rudman-etal-2023-outlier
%X Representations from large language models (LLMs) are known to be dominated by a small subset of dimensions with exceedingly high variance. Previous works have argued that although ablating these outlier dimensions in LLM representations hurts downstream performance, outlier dimensions are detrimental to the representational quality of embeddings. In this study, we investigate how fine-tuning impacts outlier dimensions and show that 1) outlier dimensions that occur in pre-training persist in fine-tuned models and 2) a single outlier dimension can complete downstream tasks with a minimal error rate. Our results suggest that outlier dimensions can encode crucial task-specific knowledge and that the value of a representation in a single outlier dimension drives downstream model decisions.
%R 10.18653/v1/2023.emnlp-main.901
%U https://aclanthology.org/2023.emnlp-main.901
%U https://doi.org/10.18653/v1/2023.emnlp-main.901
%P 14596-14605
Markdown (Informal)
[Outlier Dimensions Encode Task Specific Knowledge](https://aclanthology.org/2023.emnlp-main.901) (Rudman et al., EMNLP 2023)
ACL
- William Rudman, Catherine Chen, and Carsten Eickhoff. 2023. Outlier Dimensions Encode Task Specific Knowledge. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 14596–14605, Singapore. Association for Computational Linguistics.