{"id":"https://openalex.org/W2994635323","doi":"https://doi.org/10.48550/arxiv.2001.07384","title":"Understanding Why Neural Networks Generalize Well Through GSNR of Parameters","display_name":"Understanding Why Neural Networks Generalize Well Through GSNR of Parameters","publication_year":2020,"publication_date":"2020-01-21","ids":{"openalex":"https://openalex.org/W2994635323","doi":"https://doi.org/10.48550/arxiv.2001.07384","mag":"2994635323"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2001.07384","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2001.07384","pdf_url":"https://arxiv.org/pdf/2001.07384","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":"public-domain","license_id":"https://openalex.org/licenses/public-domain","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"volume":null,"type":"preprint","is_traditional":true,"indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2001.07384","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5100339170","display_name":"Jinlong Liu","orcid":"https://orcid.org/0000-0002-3571-0526","observed_orcids":["https://orcid.org/0000-0002-3571-0526"]},"institutions":[{"id":"https://openalex.org/I4210155230","display_name":"Samsung (China)","ror":"https://ror.org/04yt00889","country_code":"CN","type":"company","lineage":["https://openalex.org/I2250650973","https://openalex.org/I4210155230"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Liu, Jinlong","raw_affiliation_strings":["Samsung Research China -Beijing (SRC-B)"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Samsung Research China -Beijing (SRC-B)","institution_ids":["https://openalex.org/I4210155230"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5103071086","display_name":"Guoqing Jiang","orcid":"https://orcid.org/0000-0002-2289-0855","observed_orcids":["https://orcid.org/0000-0002-2289-0855"]},"institutions":[{"id":"https://openalex.org/I4210155230","display_name":"Samsung (China)","ror":"https://ror.org/04yt00889","country_code":"CN","type":"company","lineage":["https://openalex.org/I2250650973","https://openalex.org/I4210155230"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Jiang, Guoqing","raw_affiliation_strings":["Samsung Research China -Beijing (SRC-B)"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Samsung Research China -Beijing (SRC-B)","institution_ids":["https://openalex.org/I4210155230"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5007105124","display_name":"Yunzhi Bai","orcid":"https://orcid.org/0009-0001-0889-0154","observed_orcids":["https://orcid.org/0009-0001-0889-0154"]},"institutions":[{"id":"https://openalex.org/I4210155230","display_name":"Samsung (China)","ror":"https://ror.org/04yt00889","country_code":"CN","type":"company","lineage":["https://openalex.org/I2250650973","https://openalex.org/I4210155230"]},{"id":"https://openalex.org/I63966007","display_name":"Massachusetts Institute of Technology","ror":"https://ror.org/042nb2s44","country_code":"US","type":"education","lineage":["https://openalex.org/I63966007"]}],"countries":["CN","US"],"is_corresponding":false,"raw_author_name":"Bai, Yunzhi","raw_affiliation_strings":["Samsung Research China -Beijing (SRC-B)","Massachusetts Institute of Technology, Cambridge, United States"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Samsung Research China -Beijing (SRC-B)","institution_ids":["https://openalex.org/I4210155230"]},{"raw_affiliation_string":"Massachusetts Institute of Technology, Cambridge, United States","institution_ids":["https://openalex.org/I63966007"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5100443189","display_name":"Ting Chen","orcid":"https://orcid.org/0000-0002-3228-9166","observed_orcids":["https://orcid.org/0000-0002-3228-9166"]},"institutions":[{"id":"https://openalex.org/I4210155230","display_name":"Samsung (China)","ror":"https://ror.org/04yt00889","country_code":"CN","type":"company","lineage":["https://openalex.org/I2250650973","https://openalex.org/I4210155230"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Chen, Ting","raw_affiliation_strings":["Samsung Research China -Beijing (SRC-B)"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Samsung Research China -Beijing (SRC-B)","institution_ids":["https://openalex.org/I4210155230"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5066451569","display_name":"Huayan Wang","orcid":"https://orcid.org/0000-0002-0168-0306","observed_orcids":["https://orcid.org/0000-0002-0168-0306"]},"institutions":[{"id":"https://openalex.org/I4210155230","display_name":"Samsung (China)","ror":"https://ror.org/04yt00889","country_code":"CN","type":"company","lineage":["https://openalex.org/I2250650973","https://openalex.org/I4210155230"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Wang, Huayan","raw_affiliation_strings":["Samsung Research China -Beijing (SRC-B)"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Samsung Research China -Beijing (SRC-B)","institution_ids":["https://openalex.org/I4210155230"]}]}],"institutions":[],"countries_distinct_count":2,"institutions_distinct_count":2,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":true,"cited_by_count":43,"citation_normalized_percentile":null,"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":null,"last_page":null},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T11612","display_name":"Stochastic Gradient Optimization Techniques","score":0.8931000232696533,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},"topics":[{"id":"https://openalex.org/T11612","display_name":"Stochastic Gradient Optimization Techniques","score":0.8931000232696533,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T10320","display_name":"Neural Networks and Applications","score":0.07540000230073929,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T11307","display_name":"Domain Adaptation and Few-Shot Learning","score":0.0142000000923872,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}}],"keywords":[{"id":"https://openalex.org/keywords/generalization-gap","display_name":"generalization gap","score":0.8169999718666077},{"id":"https://openalex.org/keywords/deep-neural-network","display_name":"deep neural network","score":0.8169999718666077},{"id":"https://openalex.org/keywords/gradient-descent-dynamics","display_name":"gradient descent dynamics","score":0.6499999761581421},{"id":"https://openalex.org/keywords/generalization","display_name":"generalization","score":0.3569999933242798},{"id":"https://openalex.org/keywords/model-parameters","display_name":"model parameters","score":0.5440000295639038},{"id":"https://openalex.org/keywords/training-dynamics","display_name":"training dynamics","score":0.49000000953674316}],"concepts":[{"id":"https://openalex.org/C177148314","wikidata":"https://www.wikidata.org/wiki/Q170084","display_name":"Generalization","level":2,"score":0.851386547088623},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6701323986053467},{"id":"https://openalex.org/C206688291","wikidata":"https://www.wikidata.org/wiki/Q7617819","display_name":"Stochastic gradient descent","level":3,"score":0.6434566378593445},{"id":"https://openalex.org/C196083921","wikidata":"https://www.wikidata.org/wiki/Q7915758","display_name":"Variance (accounting)","level":2,"score":0.6218739748001099},{"id":"https://openalex.org/C153258448","wikidata":"https://www.wikidata.org/wiki/Q1199743","display_name":"Gradient descent","level":3,"score":0.6109393239021301},{"id":"https://openalex.org/C50644808","wikidata":"https://www.wikidata.org/wiki/Q192776","display_name":"Artificial neural network","level":2,"score":0.5906636714935303},{"id":"https://openalex.org/C98045186","wikidata":"https://www.wikidata.org/wiki/Q205663","display_name":"Process (computing)","level":2,"score":0.5230608582496643},{"id":"https://openalex.org/C12713177","wikidata":"https://www.wikidata.org/wiki/Q1900281","display_name":"Perspective (graphical)","level":2,"score":0.5182768702507019},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.5028049945831299},{"id":"https://openalex.org/C2984842247","wikidata":"https://www.wikidata.org/wiki/Q197536","display_name":"Deep neural networks","level":3,"score":0.48567995429039},{"id":"https://openalex.org/C99498987","wikidata":"https://www.wikidata.org/wiki/Q2210247","display_name":"Noise (video)","level":3,"score":0.4552038013935089},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.44669222831726074},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.40970736742019653},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.22849693894386292},{"id":"https://openalex.org/C115961682","wikidata":"https://www.wikidata.org/wiki/Q860623","display_name":"Image (mathematics)","level":2,"score":0.0},{"id":"https://openalex.org/C121955636","wikidata":"https://www.wikidata.org/wiki/Q4116214","display_name":"Accounting","level":1,"score":0.0},{"id":"https://openalex.org/C134306372","wikidata":"https://www.wikidata.org/wiki/Q7754","display_name":"Mathematical analysis","level":1,"score":0.0},{"id":"https://openalex.org/C111919701","wikidata":"https://www.wikidata.org/wiki/Q9135","display_name":"Operating system","level":1,"score":0.0},{"id":"https://openalex.org/C144133560","wikidata":"https://www.wikidata.org/wiki/Q4830453","display_name":"Business","level":0,"score":0.0},{"id":"https://openalex.org/C38652104","wikidata":"https://www.wikidata.org/wiki/Q3510521","display_name":"Computer security","level":1,"score":0.0}],"mesh":[],"locations_count":3,"locations":[{"id":"pmh:oai:arXiv.org:2001.07384","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2001.07384","pdf_url":"https://arxiv.org/pdf/2001.07384","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":"public-domain","license_id":"https://openalex.org/licenses/public-domain","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2001.07384","doi":"https://doi.org/10.48550/arxiv.2001.07384","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2001.07384","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":"public-domain","license_id":"https://openalex.org/licenses/public-domain","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"},{"id":"mag:2994635323","doi":null,"is_oa":false,"landing_page_url":"https://www.openreview.net/pdf?id=HyevIJStwH","pdf_url":null,"source":{"id":"https://openalex.org/S4306419637","display_name":"International Conference on Learning Representations","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"conference"},"license":null,"license_id":null,"version":null,"is_accepted":false,"is_published":false,"raw_source_name":"International Conference on Learning Representations","raw_type":null}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2001.07384","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2001.07384","pdf_url":"https://arxiv.org/pdf/2001.07384","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":"public-domain","license_id":"https://openalex.org/licenses/public-domain","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"sustainable_development_goals":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W2994635323.pdf","grobid_xml":"https://content.openalex.org/works/W2994635323.grobid-xml"},"referenced_works_count":17,"referenced_works":["https://openalex.org/W6908809","https://openalex.org/W1522301498","https://openalex.org/W1810943226","https://openalex.org/W1811750039","https://openalex.org/W2523060838","https://openalex.org/W2579923771","https://openalex.org/W2604117713","https://openalex.org/W2763894180","https://openalex.org/W2766196653","https://openalex.org/W2785749378","https://openalex.org/W2785811906","https://openalex.org/W2786622092","https://openalex.org/W2914704901","https://openalex.org/W2950220847","https://openalex.org/W2963655672","https://openalex.org/W2964121744","https://openalex.org/W4310348270"],"related_works":["https://openalex.org/W4221158304","https://openalex.org/W4417137028","https://openalex.org/W7133204659","https://openalex.org/W4416074406","https://openalex.org/W2996822312","https://openalex.org/W4225693694","https://openalex.org/W2927724204","https://openalex.org/W4287244223","https://openalex.org/W7163898681","https://openalex.org/W2995185977"],"abstract_inverted_index":{"As":[0],"deep":[1],"neural":[2],"networks":[3],"(DNNs)":[4],"achieve":[5],"tremendous":[6],"success":[7],"across":[8],"many":[9,17],"application":[10],"domains,":[11],"researchers":[12],"tried":[13],"to":[14,39,98,136],"explore":[15],"in":[16],"aspects":[18],"on":[19,32,72],"why":[20],"they":[21],"generalize":[22],"well.":[23],"In":[24],"this":[25],"paper,":[26],"we":[27,75,103],"provide":[28],"a":[29,53,77],"novel":[30],"perspective":[31],"these":[33],"issues":[34],"using":[35],"the":[36,58,68,85,118,134],"gradient":[37,119],"signal":[38],"noise":[40],"ratio":[41,59],"(GSNR)":[42],"of":[43,48,52,109,123],"parameters":[44],"during":[45,94,129],"training":[46,95],"process":[47,96],"DNNs.":[49],"The":[50],"GSNR":[51,83,93,128],"parameter":[54],"is":[55,132],"defined":[56],"as":[57],"between":[60,80],"its":[61],"gradient's":[62],"squared":[63],"mean":[64],"and":[65,84],"variance,":[66],"over":[67],"data":[69],"distribution.":[70],"Based":[71],"several":[73],"approximations,":[74],"establish":[76],"quantitative":[78],"relationship":[79,89],"model":[81],"parameters'":[82],"generalization":[86,100,139],"gap.":[87],"This":[88],"indicates":[90],"that":[91,108],"larger":[92],"leads":[97],"better":[99],"performance.":[101],"Moreover,":[102],"show":[104],"that,":[105],"different":[106],"from":[107],"shallow":[110],"models":[111],"(e.g.":[112],"logistic":[113],"regression,":[114],"support":[115],"vector":[116],"machines),":[117],"descent":[120],"optimization":[121],"dynamics":[122],"DNNs":[124],"naturally":[125],"produces":[126],"large":[127],"training,":[130],"which":[131],"probably":[133],"key":[135],"DNNs'":[137],"remarkable":[138],"ability.":[140]},"counts_by_year":[{"year":2025,"cited_by_count":8},{"year":2024,"cited_by_count":4},{"year":2023,"cited_by_count":7},{"year":2022,"cited_by_count":6},{"year":2021,"cited_by_count":12},{"year":2020,"cited_by_count":6}],"updated_date":"2026-11-10T11:39:43.021295","created_date":"2019-12-26T00:00:00"}