Beyond Distribution Shift: Spurious Features Through the Lens of Training Dynamics

Deep Neural Networks (DNNs) are prone to learning spurious features that correlate with the label during training but are irrelevant to the learning problem. This hurts model generalization and poses problems when deploying them in safety-critical applications. This paper aims to better understand t...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:arXiv.org 2023-10
Hauptverfasser: Murali, Nihal, Puli, Aahlad, Yu, Ke, Ranganath, Rajesh, Kayhan Batmanghelich
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 Murali, Nihal
Puli, Aahlad
Yu, Ke
Ranganath, Rajesh
Kayhan Batmanghelich
description Deep Neural Networks (DNNs) are prone to learning spurious features that correlate with the label during training but are irrelevant to the learning problem. This hurts model generalization and poses problems when deploying them in safety-critical applications. This paper aims to better understand the effects of spurious features through the lens of the learning dynamics of the internal neurons during the training process. We make the following observations: (1) While previous works highlight the harmful effects of spurious features on the generalization ability of DNNs, we emphasize that not all spurious features are harmful. Spurious features can be "benign" or "harmful" depending on whether they are "harder" or "easier" to learn than the core features for a given model. This definition is model and dataset-dependent. (2) We build upon this premise and use instance difficulty methods (like Prediction Depth (Baldock et al., 2021)) to quantify "easiness" for a given model and to identify this behavior during the training phase. (3) We empirically show that the harmful spurious features can be detected by observing the learning dynamics of the DNN's early layers. In other words, easy features learned by the initial layers of a DNN early during the training can (potentially) hurt model generalization. We verify our claims on medical and vision datasets, both simulated and real, and justify the empirical success of our hypothesis by showing the theoretical connections between Prediction Depth and information-theoretic concepts like V-usable information (Ethayarajh et al., 2021). Lastly, our experiments show that monitoring only accuracy during training (as is common in machine learning pipelines) is insufficient to detect spurious features. We, therefore, highlight the need for monitoring early training dynamics using suitable instance difficulty metrics.
format Article
fullrecord <record><control><sourceid>proquest</sourceid><recordid>TN_cdi_proquest_journals_2778492674</recordid><sourceformat>XML</sourceformat><sourcesystem>PC</sourcesystem><sourcerecordid>2778492674</sourcerecordid><originalsourceid>FETCH-proquest_journals_27784926743</originalsourceid><addsrcrecordid>eNqNjrEOgjAUABsTE4nyDy9xJsECgo6KxMFJ2UnVQh_RFvvagb-XwQ9wuuFuuBkLeJJsoiLlfMFCoj6OY77NeZYlAbse5Gj0E0okZ_HuHRoNN4Wt28Nt8BaNJ6ikcN5KglpZ4zsFTkm4SE1gWqitQI26g3LU4o0PWrF5K14kwx-XbF2d6uM5Gqz5eEmu6Y23elINz_Mi3U0zafJf9QXzRT_G</addsrcrecordid><sourcetype>Aggregation Database</sourcetype><iscdi>true</iscdi><recordtype>article</recordtype><pqid>2778492674</pqid></control><display><type>article</type><title>Beyond Distribution Shift: Spurious Features Through the Lens of Training Dynamics</title><source>Free E- Journals</source><creator>Murali, Nihal ; Puli, Aahlad ; Yu, Ke ; Ranganath, Rajesh ; Kayhan Batmanghelich</creator><creatorcontrib>Murali, Nihal ; Puli, Aahlad ; Yu, Ke ; Ranganath, Rajesh ; Kayhan Batmanghelich</creatorcontrib><description>Deep Neural Networks (DNNs) are prone to learning spurious features that correlate with the label during training but are irrelevant to the learning problem. This hurts model generalization and poses problems when deploying them in safety-critical applications. This paper aims to better understand the effects of spurious features through the lens of the learning dynamics of the internal neurons during the training process. We make the following observations: (1) While previous works highlight the harmful effects of spurious features on the generalization ability of DNNs, we emphasize that not all spurious features are harmful. Spurious features can be "benign" or "harmful" depending on whether they are "harder" or "easier" to learn than the core features for a given model. This definition is model and dataset-dependent. (2) We build upon this premise and use instance difficulty methods (like Prediction Depth (Baldock et al., 2021)) to quantify "easiness" for a given model and to identify this behavior during the training phase. (3) We empirically show that the harmful spurious features can be detected by observing the learning dynamics of the DNN's early layers. In other words, easy features learned by the initial layers of a DNN early during the training can (potentially) hurt model generalization. We verify our claims on medical and vision datasets, both simulated and real, and justify the empirical success of our hypothesis by showing the theoretical connections between Prediction Depth and information-theoretic concepts like V-usable information (Ethayarajh et al., 2021). Lastly, our experiments show that monitoring only accuracy during training (as is common in machine learning pipelines) is insufficient to detect spurious features. We, therefore, highlight the need for monitoring early training dynamics using suitable instance difficulty metrics.</description><identifier>EISSN: 2331-8422</identifier><language>eng</language><publisher>Ithaca: Cornell University Library, arXiv.org</publisher><subject>Artificial neural networks ; Computer architecture ; Damage patterns ; Dynamics ; Information theory ; Lenses ; Machine learning ; Medical imaging ; Monitoring ; Safety critical ; Training</subject><ispartof>arXiv.org, 2023-10</ispartof><rights>2023. 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>776,780</link.rule.ids></links><search><creatorcontrib>Murali, Nihal</creatorcontrib><creatorcontrib>Puli, Aahlad</creatorcontrib><creatorcontrib>Yu, Ke</creatorcontrib><creatorcontrib>Ranganath, Rajesh</creatorcontrib><creatorcontrib>Kayhan Batmanghelich</creatorcontrib><title>Beyond Distribution Shift: Spurious Features Through the Lens of Training Dynamics</title><title>arXiv.org</title><description>Deep Neural Networks (DNNs) are prone to learning spurious features that correlate with the label during training but are irrelevant to the learning problem. This hurts model generalization and poses problems when deploying them in safety-critical applications. This paper aims to better understand the effects of spurious features through the lens of the learning dynamics of the internal neurons during the training process. We make the following observations: (1) While previous works highlight the harmful effects of spurious features on the generalization ability of DNNs, we emphasize that not all spurious features are harmful. Spurious features can be "benign" or "harmful" depending on whether they are "harder" or "easier" to learn than the core features for a given model. This definition is model and dataset-dependent. (2) We build upon this premise and use instance difficulty methods (like Prediction Depth (Baldock et al., 2021)) to quantify "easiness" for a given model and to identify this behavior during the training phase. (3) We empirically show that the harmful spurious features can be detected by observing the learning dynamics of the DNN's early layers. In other words, easy features learned by the initial layers of a DNN early during the training can (potentially) hurt model generalization. We verify our claims on medical and vision datasets, both simulated and real, and justify the empirical success of our hypothesis by showing the theoretical connections between Prediction Depth and information-theoretic concepts like V-usable information (Ethayarajh et al., 2021). Lastly, our experiments show that monitoring only accuracy during training (as is common in machine learning pipelines) is insufficient to detect spurious features. We, therefore, highlight the need for monitoring early training dynamics using suitable instance difficulty metrics.</description><subject>Artificial neural networks</subject><subject>Computer architecture</subject><subject>Damage patterns</subject><subject>Dynamics</subject><subject>Information theory</subject><subject>Lenses</subject><subject>Machine learning</subject><subject>Medical imaging</subject><subject>Monitoring</subject><subject>Safety critical</subject><subject>Training</subject><issn>2331-8422</issn><fulltext>true</fulltext><rsrctype>article</rsrctype><creationdate>2023</creationdate><recordtype>article</recordtype><sourceid>BENPR</sourceid><recordid>eNqNjrEOgjAUABsTE4nyDy9xJsECgo6KxMFJ2UnVQh_RFvvagb-XwQ9wuuFuuBkLeJJsoiLlfMFCoj6OY77NeZYlAbse5Gj0E0okZ_HuHRoNN4Wt28Nt8BaNJ6ikcN5KglpZ4zsFTkm4SE1gWqitQI26g3LU4o0PWrF5K14kwx-XbF2d6uM5Gqz5eEmu6Y23elINz_Mi3U0zafJf9QXzRT_G</recordid><startdate>20231014</startdate><enddate>20231014</enddate><creator>Murali, Nihal</creator><creator>Puli, Aahlad</creator><creator>Yu, Ke</creator><creator>Ranganath, Rajesh</creator><creator>Kayhan Batmanghelich</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>20231014</creationdate><title>Beyond Distribution Shift: Spurious Features Through the Lens of Training Dynamics</title><author>Murali, Nihal ; Puli, Aahlad ; Yu, Ke ; Ranganath, Rajesh ; Kayhan Batmanghelich</author></sort><facets><frbrtype>5</frbrtype><frbrgroupid>cdi_FETCH-proquest_journals_27784926743</frbrgroupid><rsrctype>articles</rsrctype><prefilter>articles</prefilter><language>eng</language><creationdate>2023</creationdate><topic>Artificial neural networks</topic><topic>Computer architecture</topic><topic>Damage patterns</topic><topic>Dynamics</topic><topic>Information theory</topic><topic>Lenses</topic><topic>Machine learning</topic><topic>Medical imaging</topic><topic>Monitoring</topic><topic>Safety critical</topic><topic>Training</topic><toplevel>online_resources</toplevel><creatorcontrib>Murali, Nihal</creatorcontrib><creatorcontrib>Puli, Aahlad</creatorcontrib><creatorcontrib>Yu, Ke</creatorcontrib><creatorcontrib>Ranganath, Rajesh</creatorcontrib><creatorcontrib>Kayhan Batmanghelich</creatorcontrib><collection>ProQuest SciTech Collection</collection><collection>ProQuest Technology Collection</collection><collection>Materials Science &amp; Engineering Collection</collection><collection>ProQuest Central (Alumni Edition)</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 Korea</collection><collection>SciTech Premium Collection</collection><collection>ProQuest Engineering Collection</collection><collection>Engineering Database</collection><collection>Publicly Available Content Database</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>Murali, Nihal</au><au>Puli, Aahlad</au><au>Yu, Ke</au><au>Ranganath, Rajesh</au><au>Kayhan Batmanghelich</au><format>book</format><genre>document</genre><ristype>GEN</ristype><atitle>Beyond Distribution Shift: Spurious Features Through the Lens of Training Dynamics</atitle><jtitle>arXiv.org</jtitle><date>2023-10-14</date><risdate>2023</risdate><eissn>2331-8422</eissn><abstract>Deep Neural Networks (DNNs) are prone to learning spurious features that correlate with the label during training but are irrelevant to the learning problem. This hurts model generalization and poses problems when deploying them in safety-critical applications. This paper aims to better understand the effects of spurious features through the lens of the learning dynamics of the internal neurons during the training process. We make the following observations: (1) While previous works highlight the harmful effects of spurious features on the generalization ability of DNNs, we emphasize that not all spurious features are harmful. Spurious features can be "benign" or "harmful" depending on whether they are "harder" or "easier" to learn than the core features for a given model. This definition is model and dataset-dependent. (2) We build upon this premise and use instance difficulty methods (like Prediction Depth (Baldock et al., 2021)) to quantify "easiness" for a given model and to identify this behavior during the training phase. (3) We empirically show that the harmful spurious features can be detected by observing the learning dynamics of the DNN's early layers. In other words, easy features learned by the initial layers of a DNN early during the training can (potentially) hurt model generalization. We verify our claims on medical and vision datasets, both simulated and real, and justify the empirical success of our hypothesis by showing the theoretical connections between Prediction Depth and information-theoretic concepts like V-usable information (Ethayarajh et al., 2021). Lastly, our experiments show that monitoring only accuracy during training (as is common in machine learning pipelines) is insufficient to detect spurious features. We, therefore, highlight the need for monitoring early training dynamics using suitable instance difficulty metrics.</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, 2023-10
issn 2331-8422
language eng
recordid cdi_proquest_journals_2778492674
source Free E- Journals
subjects Artificial neural networks
Computer architecture
Damage patterns
Dynamics
Information theory
Lenses
Machine learning
Medical imaging
Monitoring
Safety critical
Training
title Beyond Distribution Shift: Spurious Features Through the Lens of Training Dynamics
url https://sfx.bib-bvb.de/sfx_tum?ctx_ver=Z39.88-2004&ctx_enc=info:ofi/enc:UTF-8&ctx_tim=2025-01-30T04%3A12%3A27IST&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=Beyond%20Distribution%20Shift:%20Spurious%20Features%20Through%20the%20Lens%20of%20Training%20Dynamics&rft.jtitle=arXiv.org&rft.au=Murali,%20Nihal&rft.date=2023-10-14&rft.eissn=2331-8422&rft_id=info:doi/&rft_dat=%3Cproquest%3E2778492674%3C/proquest%3E%3Curl%3E%3C/url%3E&disable_directlink=true&sfx.directlink=off&sfx.report_link=0&rft_id=info:oai/&rft_pqid=2778492674&rft_id=info:pmid/&rfr_iscdi=true