Deep Learning with Zero Initialization: Revisiting Symmetry Breaking and Gradient Flow

Deep Learning with Zero Initialization: Revisiting Symmetry Breaking and Gradient Flow | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Deep Learning with Zero Initialization: Revisiting Symmetry Breaking and Gradient Flow Jongwoo Seo, Wuhyun Koh This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-4890533/v3 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 16 Mar, 2026 Read the published version in Neural Processing Letters → Version 3 posted 6 You are reading this latest preprint version Show more versions Abstract For decades, the artificial intelligence (AI) community has believed that zero initialization is ineffective for neural networks. Our study challenges this misconception by introducing a method that enables successful learning even when all weights and biases are initialized to zero. Beyond this method, we also examine mixed initialization schemes in which zero and random initialization coexist across different layers or parameters, showing that learning remains effective even under such partially randomized settings. Experiments on MNIST, CIFAR-10, CIFAR-100, and Tiny ImageNet using multilayer perceptrons (MLPs), convolutional neural networks (CNNs), residual networks (ResNets), vision transformers (ViTs), and multilayer perceptron mixers (MLP-Mixers) show that zero initialization can match or even surpass random initialization in certain scenarios, particularly with MLPs and CNNs. Notably, MLP-Mixers retained comparable performance despite having no randomly initialized parameters. These findings position random initialization as a special case of zero-centered symmetry breaking and refute the longstanding belief that zero initialization inherently degrades neural network performance, opening new possibilities for neural network training. To systematize these insights, we propose the "Seo Integrated Zero Initialization: Foundational Scheme (SIZIFS)" — a unified conceptual structure that categorizes artificial neural network initialization strategies into weight-level, node-level, and context-dependent types. Implementation code is publicly available at: https://github.com/sjw007s/Deep-Learning-with-Zero-Initialization-Revisiting-Symmetry-Breaking-and-Gradient-Flow . Artificial Neural Network Neural Network Initialization Weight Initialization Zero Initialization Full Text Additional Declarations No competing interests reported. Cite Share Download PDF Status: Published Journal Publication published 16 Mar, 2026 Read the published version in Neural Processing Letters → Version 3 posted Editorial decision: Revision requested 01 Feb, 2026 Reviews received at journal 10 Nov, 2025 Reviewers agreed at journal 10 Nov, 2025 Reviewers invited by journal 09 Nov, 2025 Submission checks completed at journal 06 Nov, 2025 First submitted to journal 03 Nov, 2025 You are reading this latest preprint version Show more versions Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-4890533","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[{"code":1,"date":"2024-08-15 17:17:48","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}},{"code":2,"date":"2025-04-07 15:39:36","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":569175641,"identity":"44faae53-f022-4995-b71f-ed5936fada51","order_by":0,"name":"Jongwoo Seo","email":"","orcid":"","institution":"Institute for Basic Science","correspondingAuthor":false,"prefix":"","firstName":"Jongwoo","middleName":"","lastName":"Seo","suffix":""},{"id":569175642,"identity":"b94d2d99-b62e-4615-b2e1-b85404626290","order_by":1,"name":"Wuhyun Koh","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAyUlEQVRIie3PMQrCMBTG8Sddn3VNCPQMlUIXi14loZBRBI/QtdC14CUEwckhEKhL3Qt1dg44Omi76WLi5pD/8IbAD74A+Hz/GBmPAj4L8O3BidDiZxIrVxL35fy+abN1cp5qMzllQHfKQq5twupOblMdSgI3CSzk30naSc7QaHHUmA4LNURoGTaQ/DGSQ4GJAfV0IXnDsNNiH2BMQClgNrIayAJbKerxL1zlSEsLobUoemwyUVUXbYxaRqS1kM84gG2Wz+fz+Vx6AQ08QhEcDGlxAAAAAElFTkSuQmCC","orcid":"","institution":"Institute for Basic Science","correspondingAuthor":true,"prefix":"","firstName":"Wuhyun","middleName":"","lastName":"Koh","suffix":""}],"badges":[],"createdAt":"2024-08-10 07:33:20","currentVersionCode":3,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-4890533/v3","doiUrl":"https://doi.org/10.21203/rs.3.rs-4890533/v3","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1007/s11063-026-11846-1","type":"published","date":"2026-03-16T15:58:08+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":105223505,"identity":"202e8f16-7d18-478a-afd9-c2d3a4aefe60","added_by":"auto","created_at":"2026-03-23 16:07:52","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":5858305,"visible":true,"origin":"","legend":"","description":"","filename":"NeuralProcessingLettersSecondRevision.pdf","url":"https://assets-eu.researchsquare.com/files/rs-4890533/v3_covered_429eeb20-5289-4a74-9799-4da2dc934245.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Deep Learning with Zero Initialization: Revisiting Symmetry Breaking and Gradient Flow","fulltext":[],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":false,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":true,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":true,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"neural-processing-letters","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"nepl","sideBox":"Learn more about [Neural Processing Letters](http://link.springer.com/journal/11063)","snPcode":"11063","submissionUrl":"https://submission.nature.com/new-submission/11063/3","title":"Neural Processing Letters","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"Artificial Neural Network, Neural Network Initialization, Weight Initialization, Zero Initialization","lastPublishedDoi":"10.21203/rs.3.rs-4890533/v3","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-4890533/v3","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"For decades, the artificial intelligence (AI) community has believed that zero initialization is ineffective for neural networks. Our study challenges this misconception by introducing a method that enables successful learning even when all weights and biases are initialized to zero. Beyond this method, we also examine mixed initialization schemes in which zero and random initialization coexist across different layers or parameters, showing that learning remains effective even under such partially randomized settings. Experiments on MNIST, CIFAR-10, CIFAR-100, and Tiny ImageNet using multilayer perceptrons (MLPs), convolutional neural networks (CNNs), residual networks (ResNets), vision transformers (ViTs), and multilayer perceptron mixers (MLP-Mixers) show that zero initialization can match or even surpass random initialization in certain scenarios, particularly with MLPs and CNNs. Notably, MLP-Mixers retained comparable performance despite having no randomly initialized parameters. These findings position random initialization as a special case of zero-centered symmetry breaking and refute the longstanding belief that zero initialization inherently degrades neural network performance, opening new possibilities for neural network training. To systematize these insights, we propose the \"Seo Integrated Zero Initialization: Foundational Scheme (SIZIFS)\" — a unified conceptual structure that categorizes artificial neural network initialization strategies into weight-level, node-level, and context-dependent types. Implementation code is publicly available at: https://github.com/sjw007s/Deep-Learning-with-Zero-Initialization-Revisiting-Symmetry-Breaking-and-Gradient-Flow.","manuscriptTitle":"Deep Learning with Zero Initialization: Revisiting Symmetry Breaking and Gradient Flow","msid":"","msnumber":"","nonDraftVersions":[{"code":3,"date":"2025-12-17 13:58:03","doi":"10.21203/rs.3.rs-4890533/v3","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-02-01T12:24:02+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-11-11T03:03:45+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"65321180903136999498693580652952973596","date":"2025-11-11T02:41:32+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-11-09T11:13:15+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-11-06T13:46:04+00:00","index":"","fulltext":""},{"type":"submitted","content":"Neural Processing Letters","date":"2025-11-03T12:05:39+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"neural-processing-letters","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"nepl","sideBox":"Learn more about [Neural Processing Letters](http://link.springer.com/journal/11063)","snPcode":"11063","submissionUrl":"https://submission.nature.com/new-submission/11063/3","title":"Neural Processing Letters","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}},{"code":"","date":"2025-10-28 21:51:46","doi":"","editorialEvents":[{"type":"decision","content":"Revision requested","date":"2025-10-28T21:51:46+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-10-22T06:04:59+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-10-17T20:36:11+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"65321180903136999498693580652952973596","date":"2025-09-25T01:29:29+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"187340456485813991647810255405102174396","date":"2025-09-20T18:45:28+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-09-19T14:43:03+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-09-19T01:04:41+00:00","index":"","fulltext":""},{"type":"submitted","content":"Neural Processing Letters","date":"2025-09-16T11:05:52+00:00","index":"","fulltext":""},{"type":"notPreprinted","content":""}],"status":"timeline","journal":{"display":true,"email":"[email protected]","identity":"neural-processing-letters","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"nepl","sideBox":"Learn more about [Neural Processing Letters](http://link.springer.com/journal/11063)","snPcode":"11063","submissionUrl":"https://submission.nature.com/new-submission/11063/3","title":"Neural Processing Letters","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}},{"code":"","date":"2025-08-11 02:38:15","doi":"","editorialEvents":[{"type":"decision","content":"Revision requested","date":"2025-09-10T14:26:09+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-08-23T17:29:38+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"314398657216346232839556321951151997513","date":"2025-08-23T16:28:58+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-08-19T09:26:56+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"187340456485813991647810255405102174396","date":"2025-08-18T21:19:55+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"65321180903136999498693580652952973596","date":"2025-08-18T12:13:57+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-08-18T11:59:43+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-08-14T15:01:19+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-08-14T15:00:43+00:00","index":"","fulltext":""},{"type":"submitted","content":"Neural Processing Letters","date":"2025-08-11T02:34:06+00:00","index":"","fulltext":""}],"status":"private","journal":{"display":true,"email":"[email protected]","identity":"neural-processing-letters","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"nepl","sideBox":"Learn more about [Neural Processing Letters](http://link.springer.com/journal/11063)","snPcode":"11063","submissionUrl":"https://submission.nature.com/new-submission/11063/3","title":"Neural Processing Letters","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"5e0c3b92-5339-4fa1-b4b2-1203173ab308","owner":[],"postedDate":"December 17th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2026-03-23T16:04:31+00:00","versionOfRecord":{"articleIdentity":"rs-4890533","link":"https://doi.org/10.1007/s11063-026-11846-1","journal":{"identity":"neural-processing-letters","isVorOnly":false,"title":"Neural Processing Letters"},"publishedOn":"2026-03-16 15:58:08","publishedOnDateReadable":"March 16th, 2026"},"versionCreatedAt":"2025-12-17 13:58:03","video":"","vorDoi":"10.1007/s11063-026-11846-1","vorDoiUrl":"https://doi.org/10.1007/s11063-026-11846-1","workflowStages":[]},"version":"v3","identity":"rs-4890533","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-4890533","identity":"rs-4890533","version":["v3"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}