Unveiling and Controlling Anomalous Attention Distribution in Transformers

With the advent of large models based on the Transformer architecture, researchers have observed an anomalous phenomenon in the Attention mechanism--there is a very high attention on the first element, which is prevalent across Transformer-based models. It is crucial to understand it for the develop...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:arXiv.org 2024-07
Hauptverfasser: Yan, Ruiqing, Du, Xingbo, Deng, Haoyu, Zheng, Linghan, Sun, Qiuzhuang, Hu, Jifang, Shao, Yuhang, Jiang, Penghao, Jiang, Jinrong, Zhao, Lian
Format: Artikel
Sprache:eng
Schlagworte:
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
container_end_page
container_issue
container_start_page
container_title arXiv.org
container_volume
creator Yan, Ruiqing
Du, Xingbo
Deng, Haoyu
Zheng, Linghan
Sun, Qiuzhuang
Hu, Jifang
Shao, Yuhang
Jiang, Penghao
Jiang, Jinrong
Zhao, Lian
description With the advent of large models based on the Transformer architecture, researchers have observed an anomalous phenomenon in the Attention mechanism--there is a very high attention on the first element, which is prevalent across Transformer-based models. It is crucial to understand it for the development of techniques focusing on attention distribution, such as Key-Value (KV) Cache compression and infinite extrapolation; however, the latent cause leaves to be unknown. In this paper, we analyze such a phenomenon from the perspective of waiver phenomenon, which involves reducing the internal values of certain elements in the sequence, allowing them to absorb excess attention without affecting their contribution to information. In specific models, due to differences in positional encoding and attention patterns, we have found that the selection of waiver elements by the model can be categorized into two methods: positional-encoding-based and feature-distribution-within-elements-based.
format Article
fullrecord <record><control><sourceid>proquest</sourceid><recordid>TN_cdi_proquest_journals_3075790471</recordid><sourceformat>XML</sourceformat><sourcesystem>PC</sourcesystem><sourcerecordid>3075790471</sourcerecordid><originalsourceid>FETCH-proquest_journals_30757904713</originalsourceid><addsrcrecordid>eNqNikEKwjAQAIMgWLR_CHgupElr9FiqIp7rWSKmkpLuajbx_UrxAZ6GYWbGMqlUWWwrKRcsJxqEEHKjZV2rjJ0v8LbOO3hwA3feIsSAfvIGcDQeE_EmRgvRIfC9oxjcLU3igHfBAPUYRhtoxea98WTzH5dsfTx07al4BnwlS_E6YArwTVcldK13otKl-u_6ADmJPbY</addsrcrecordid><sourcetype>Aggregation Database</sourcetype><iscdi>true</iscdi><recordtype>article</recordtype><pqid>3075790471</pqid></control><display><type>article</type><title>Unveiling and Controlling Anomalous Attention Distribution in Transformers</title><source>Free E- Journals</source><creator>Yan, Ruiqing ; Du, Xingbo ; Deng, Haoyu ; Zheng, Linghan ; Sun, Qiuzhuang ; Hu, Jifang ; Shao, Yuhang ; Jiang, Penghao ; Jiang, Jinrong ; Zhao, Lian</creator><creatorcontrib>Yan, Ruiqing ; Du, Xingbo ; Deng, Haoyu ; Zheng, Linghan ; Sun, Qiuzhuang ; Hu, Jifang ; Shao, Yuhang ; Jiang, Penghao ; Jiang, Jinrong ; Zhao, Lian</creatorcontrib><description>With the advent of large models based on the Transformer architecture, researchers have observed an anomalous phenomenon in the Attention mechanism--there is a very high attention on the first element, which is prevalent across Transformer-based models. It is crucial to understand it for the development of techniques focusing on attention distribution, such as Key-Value (KV) Cache compression and infinite extrapolation; however, the latent cause leaves to be unknown. In this paper, we analyze such a phenomenon from the perspective of waiver phenomenon, which involves reducing the internal values of certain elements in the sequence, allowing them to absorb excess attention without affecting their contribution to information. In specific models, due to differences in positional encoding and attention patterns, we have found that the selection of waiver elements by the model can be categorized into two methods: positional-encoding-based and feature-distribution-within-elements-based.</description><identifier>EISSN: 2331-8422</identifier><language>eng</language><publisher>Ithaca: Cornell University Library, arXiv.org</publisher><subject>Coding</subject><ispartof>arXiv.org, 2024-07</ispartof><rights>2024. This work is published under http://creativecommons.org/licenses/by/4.0/ (the “License”). Notwithstanding the ProQuest Terms and Conditions, you may use this content in accordance with the terms of the License.</rights><oa>free_for_read</oa><woscitedreferencessubscribed>false</woscitedreferencessubscribed></display><links><openurl>$$Topenurl_article</openurl><openurlfulltext>$$Topenurlfull_article</openurlfulltext><thumbnail>$$Tsyndetics_thumb_exl</thumbnail><link.rule.ids>781,785</link.rule.ids></links><search><creatorcontrib>Yan, Ruiqing</creatorcontrib><creatorcontrib>Du, Xingbo</creatorcontrib><creatorcontrib>Deng, Haoyu</creatorcontrib><creatorcontrib>Zheng, Linghan</creatorcontrib><creatorcontrib>Sun, Qiuzhuang</creatorcontrib><creatorcontrib>Hu, Jifang</creatorcontrib><creatorcontrib>Shao, Yuhang</creatorcontrib><creatorcontrib>Jiang, Penghao</creatorcontrib><creatorcontrib>Jiang, Jinrong</creatorcontrib><creatorcontrib>Zhao, Lian</creatorcontrib><title>Unveiling and Controlling Anomalous Attention Distribution in Transformers</title><title>arXiv.org</title><description>With the advent of large models based on the Transformer architecture, researchers have observed an anomalous phenomenon in the Attention mechanism--there is a very high attention on the first element, which is prevalent across Transformer-based models. It is crucial to understand it for the development of techniques focusing on attention distribution, such as Key-Value (KV) Cache compression and infinite extrapolation; however, the latent cause leaves to be unknown. In this paper, we analyze such a phenomenon from the perspective of waiver phenomenon, which involves reducing the internal values of certain elements in the sequence, allowing them to absorb excess attention without affecting their contribution to information. In specific models, due to differences in positional encoding and attention patterns, we have found that the selection of waiver elements by the model can be categorized into two methods: positional-encoding-based and feature-distribution-within-elements-based.</description><subject>Coding</subject><issn>2331-8422</issn><fulltext>true</fulltext><rsrctype>article</rsrctype><creationdate>2024</creationdate><recordtype>article</recordtype><sourceid>ABUWG</sourceid><sourceid>AFKRA</sourceid><sourceid>AZQEC</sourceid><sourceid>BENPR</sourceid><sourceid>CCPQU</sourceid><sourceid>DWQXO</sourceid><recordid>eNqNikEKwjAQAIMgWLR_CHgupElr9FiqIp7rWSKmkpLuajbx_UrxAZ6GYWbGMqlUWWwrKRcsJxqEEHKjZV2rjJ0v8LbOO3hwA3feIsSAfvIGcDQeE_EmRgvRIfC9oxjcLU3igHfBAPUYRhtoxea98WTzH5dsfTx07al4BnwlS_E6YArwTVcldK13otKl-u_6ADmJPbY</recordid><startdate>20240703</startdate><enddate>20240703</enddate><creator>Yan, Ruiqing</creator><creator>Du, Xingbo</creator><creator>Deng, Haoyu</creator><creator>Zheng, Linghan</creator><creator>Sun, Qiuzhuang</creator><creator>Hu, Jifang</creator><creator>Shao, Yuhang</creator><creator>Jiang, Penghao</creator><creator>Jiang, Jinrong</creator><creator>Zhao, Lian</creator><general>Cornell University Library, arXiv.org</general><scope>8FE</scope><scope>8FG</scope><scope>ABJCF</scope><scope>ABUWG</scope><scope>AFKRA</scope><scope>AZQEC</scope><scope>BENPR</scope><scope>BGLVJ</scope><scope>CCPQU</scope><scope>DWQXO</scope><scope>HCIFZ</scope><scope>L6V</scope><scope>M7S</scope><scope>PIMPY</scope><scope>PQEST</scope><scope>PQQKQ</scope><scope>PQUKI</scope><scope>PRINS</scope><scope>PTHSS</scope></search><sort><creationdate>20240703</creationdate><title>Unveiling and Controlling Anomalous Attention Distribution in Transformers</title><author>Yan, Ruiqing ; Du, Xingbo ; Deng, Haoyu ; Zheng, Linghan ; Sun, Qiuzhuang ; Hu, Jifang ; Shao, Yuhang ; Jiang, Penghao ; Jiang, Jinrong ; Zhao, Lian</author></sort><facets><frbrtype>5</frbrtype><frbrgroupid>cdi_FETCH-proquest_journals_30757904713</frbrgroupid><rsrctype>articles</rsrctype><prefilter>articles</prefilter><language>eng</language><creationdate>2024</creationdate><topic>Coding</topic><toplevel>online_resources</toplevel><creatorcontrib>Yan, Ruiqing</creatorcontrib><creatorcontrib>Du, Xingbo</creatorcontrib><creatorcontrib>Deng, Haoyu</creatorcontrib><creatorcontrib>Zheng, Linghan</creatorcontrib><creatorcontrib>Sun, Qiuzhuang</creatorcontrib><creatorcontrib>Hu, Jifang</creatorcontrib><creatorcontrib>Shao, Yuhang</creatorcontrib><creatorcontrib>Jiang, Penghao</creatorcontrib><creatorcontrib>Jiang, Jinrong</creatorcontrib><creatorcontrib>Zhao, Lian</creatorcontrib><collection>ProQuest SciTech Collection</collection><collection>ProQuest Technology Collection</collection><collection>Materials Science &amp; Engineering Collection</collection><collection>ProQuest Central (Alumni)</collection><collection>ProQuest Central UK/Ireland</collection><collection>ProQuest Central Essentials</collection><collection>ProQuest Central</collection><collection>Technology Collection</collection><collection>ProQuest One Community College</collection><collection>ProQuest Central</collection><collection>SciTech Premium Collection (Proquest) (PQ_SDU_P3)</collection><collection>ProQuest Engineering Collection</collection><collection>Engineering Database</collection><collection>Publicly Available Content Database (Proquest) (PQ_SDU_P3)</collection><collection>ProQuest One Academic Eastern Edition (DO NOT USE)</collection><collection>ProQuest One Academic</collection><collection>ProQuest One Academic UKI Edition</collection><collection>ProQuest Central China</collection><collection>Engineering collection</collection></facets><delivery><delcategory>Remote Search Resource</delcategory><fulltext>fulltext</fulltext></delivery><addata><au>Yan, Ruiqing</au><au>Du, Xingbo</au><au>Deng, Haoyu</au><au>Zheng, Linghan</au><au>Sun, Qiuzhuang</au><au>Hu, Jifang</au><au>Shao, Yuhang</au><au>Jiang, Penghao</au><au>Jiang, Jinrong</au><au>Zhao, Lian</au><format>book</format><genre>document</genre><ristype>GEN</ristype><atitle>Unveiling and Controlling Anomalous Attention Distribution in Transformers</atitle><jtitle>arXiv.org</jtitle><date>2024-07-03</date><risdate>2024</risdate><eissn>2331-8422</eissn><abstract>With the advent of large models based on the Transformer architecture, researchers have observed an anomalous phenomenon in the Attention mechanism--there is a very high attention on the first element, which is prevalent across Transformer-based models. It is crucial to understand it for the development of techniques focusing on attention distribution, such as Key-Value (KV) Cache compression and infinite extrapolation; however, the latent cause leaves to be unknown. In this paper, we analyze such a phenomenon from the perspective of waiver phenomenon, which involves reducing the internal values of certain elements in the sequence, allowing them to absorb excess attention without affecting their contribution to information. In specific models, due to differences in positional encoding and attention patterns, we have found that the selection of waiver elements by the model can be categorized into two methods: positional-encoding-based and feature-distribution-within-elements-based.</abstract><cop>Ithaca</cop><pub>Cornell University Library, arXiv.org</pub><oa>free_for_read</oa></addata></record>
fulltext fulltext
identifier EISSN: 2331-8422
ispartof arXiv.org, 2024-07
issn 2331-8422
language eng
recordid cdi_proquest_journals_3075790471
source Free E- Journals
subjects Coding
title Unveiling and Controlling Anomalous Attention Distribution in Transformers
url https://sfx.bib-bvb.de/sfx_tum?ctx_ver=Z39.88-2004&ctx_enc=info:ofi/enc:UTF-8&ctx_tim=2024-12-17T19%3A42%3A35IST&url_ver=Z39.88-2004&url_ctx_fmt=infofi/fmt:kev:mtx:ctx&rfr_id=info:sid/primo.exlibrisgroup.com:primo3-Article-proquest&rft_val_fmt=info:ofi/fmt:kev:mtx:book&rft.genre=document&rft.atitle=Unveiling%20and%20Controlling%20Anomalous%20Attention%20Distribution%20in%20Transformers&rft.jtitle=arXiv.org&rft.au=Yan,%20Ruiqing&rft.date=2024-07-03&rft.eissn=2331-8422&rft_id=info:doi/&rft_dat=%3Cproquest%3E3075790471%3C/proquest%3E%3Curl%3E%3C/url%3E&disable_directlink=true&sfx.directlink=off&sfx.report_link=0&rft_id=info:oai/&rft_pqid=3075790471&rft_id=info:pmid/&rfr_iscdi=true