{"id":"https://openalex.org/W4387323784","doi":"https://doi.org/10.48550/arxiv.2310.00692","title":"A Theoretical Analysis of Noise Geometry in Stochastic Gradient Descent","display_name":"A Theoretical Analysis of Noise Geometry in Stochastic Gradient Descent","publication_year":2023,"publication_date":"2023-10-01","ids":{"openalex":"https://openalex.org/W4387323784","doi":"https://doi.org/10.48550/arxiv.2310.00692"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2310.00692","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2310.00692","pdf_url":"https://arxiv.org/pdf/2310.00692","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":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"volume":null,"type":"preprint","indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2310.00692","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5041887782","display_name":"Mingze Wang","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I20231570","display_name":"Peking University","ror":"https://ror.org/02v51f717","country_code":"CN","type":"education","lineage":["https://openalex.org/I20231570"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Wang, Mingze","raw_affiliation_strings":["School of Mathematical Sciences Peking University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"School of Mathematical Sciences Peking University","institution_ids":["https://openalex.org/I20231570"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5012324275","display_name":"Lei Wu","orcid":"https://orcid.org/0000-0001-7169-1505","observed_orcids":["https://orcid.org/0000-0001-7169-1505"]},"institutions":[{"id":"https://openalex.org/I20231570","display_name":"Peking University","ror":"https://ror.org/02v51f717","country_code":"CN","type":"education","lineage":["https://openalex.org/I20231570"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Wu, Lei","raw_affiliation_strings":["School of Mathematical Sciences Center for Machine Learning Research Peking University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"School of Mathematical Sciences Center for Machine Learning Research Peking University","institution_ids":["https://openalex.org/I20231570"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":1,"corresponding_author_ids":[],"corresponding_institution_ids":["https://openalex.org/I20231570"],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":true,"cited_by_count":0,"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":1.0,"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":1.0,"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/T12535","display_name":"Machine Learning and Data Classification","score":0.0,"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":0.0,"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/two-layer-neural-networks","display_name":"two-layer neural networks","score":0.8769999742507935},{"id":"https://openalex.org/keywords/flat-minima","display_name":"flat minima","score":0.8199999928474426},{"id":"https://openalex.org/keywords/subspace-projection","display_name":"subspace projection","score":0.6970000267028809},{"id":"https://openalex.org/keywords/cyclical-learning-rate","display_name":"cyclical learning rate","score":0.843999981880188},{"id":"https://openalex.org/keywords/loss-landscape","display_name":"loss landscape","score":0.43799999356269836}],"concepts":[{"id":"https://openalex.org/C2780841128","wikidata":"https://www.wikidata.org/wiki/Q5073781","display_name":"Characterization (materials science)","level":2,"score":0.6114596724510193},{"id":"https://openalex.org/C99498987","wikidata":"https://www.wikidata.org/wiki/Q2210247","display_name":"Noise (video)","level":3,"score":0.550133228302002},{"id":"https://openalex.org/C206688291","wikidata":"https://www.wikidata.org/wiki/Q7617819","display_name":"Stochastic gradient descent","level":3,"score":0.5154479146003723},{"id":"https://openalex.org/C2524010","wikidata":"https://www.wikidata.org/wiki/Q8087","display_name":"Geometry","level":1,"score":0.44114041328430176},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.3798556625843048},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.364629328250885},{"id":"https://openalex.org/C121864883","wikidata":"https://www.wikidata.org/wiki/Q677916","display_name":"Statistical physics","level":1,"score":0.36219072341918945},{"id":"https://openalex.org/C121332964","wikidata":"https://www.wikidata.org/wiki/Q413","display_name":"Physics","level":0,"score":0.3009207248687744},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.16753259301185608},{"id":"https://openalex.org/C120665830","wikidata":"https://www.wikidata.org/wiki/Q14620","display_name":"Optics","level":1,"score":0.10159552097320557},{"id":"https://openalex.org/C50644808","wikidata":"https://www.wikidata.org/wiki/Q192776","display_name":"Artificial neural network","level":2,"score":0.060633301734924316},{"id":"https://openalex.org/C115961682","wikidata":"https://www.wikidata.org/wiki/Q860623","display_name":"Image (mathematics)","level":2,"score":0.0}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2310.00692","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2310.00692","pdf_url":"https://arxiv.org/pdf/2310.00692","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":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2310.00692","doi":"https://doi.org/10.48550/arxiv.2310.00692","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2310.00692","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":null,"license_id":null,"version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2310.00692","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2310.00692","pdf_url":"https://arxiv.org/pdf/2310.00692","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":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"sustainable_development_goals":[],"study_designs":[],"awards":[],"funders":[{"id":"https://openalex.org/F4320324787","display_name":"Peking University","ror":"https://ror.org/02v51f717"}],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4387323784.pdf","grobid_xml":"https://content.openalex.org/works/W4387323784.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W7133218258","https://openalex.org/W2804386825","https://openalex.org/W7124676036","https://openalex.org/W3166125248","https://openalex.org/W3004700632","https://openalex.org/W3034840375","https://openalex.org/W2790626470","https://openalex.org/W3213489681","https://openalex.org/W3118431319","https://openalex.org/W4414581403"],"abstract_inverted_index":{"In":[0],"this":[1,130],"paper,":[2],"we":[3,89],"provide":[4],"a":[5,17,91],"theoretical":[6,148],"study":[7],"of":[8,26,75,78,84,94],"noise":[9,20,37,86],"geometry":[10,25,87],"for":[11,53],"minibatch":[12],"stochastic":[13],"gradient":[14,108],"descent":[15,109],"(SGD),":[16],"phenomenon":[18],"where":[19],"aligns":[21],"favorably":[22],"with":[23],"the":[24,39,47,66,76,82,95,114],"local":[27],"landscape.":[28],"We":[29,50,104],"propose":[30],"two":[31],"metrics,":[32,65],"derived":[33],"from":[34,101,121],"analyzing":[35],"how":[36],"influences":[38],"loss":[40],"and":[41,57,124],"subspace":[42],"projection":[43],"dynamics,":[44],"to":[45,119,133,145],"quantify":[46],"alignment":[48,67],"strength.":[49],"show":[51],"that":[52,106],"(over-parameterized)":[54],"linear":[55],"models":[56],"two-layer":[58],"nonlinear":[59],"networks,":[60],"when":[61],"measured":[62],"by":[63,97],"these":[64],"can":[68,128],"be":[69],"provably":[70],"guaranteed":[71],"under":[72],"conditions":[73],"independent":[74],"degree":[77],"over-parameterization.":[79],"To":[80],"showcase":[81],"utility":[83],"our":[85,147],"characterizations,":[88],"present":[90],"refined":[92],"analysis":[93],"mechanism":[96],"which":[98,111],"SGD":[99,117,131],"escapes":[100,112],"sharp":[102],"minima.":[103],"reveal":[105],"unlike":[107],"(GD),":[110],"along":[113],"sharpest":[115],"directions,":[116],"tends":[118],"escape":[120],"flatter":[122,138],"directions":[123],"cyclical":[125],"learning":[126],"rates":[127],"exploit":[129],"characteristic":[132],"navigate":[134],"more":[135],"effectively":[136],"towards":[137],"regions.":[139],"Lastly,":[140],"extensive":[141],"experiments":[142],"are":[143],"provided":[144],"support":[146],"findings.":[149]},"counts_by_year":[],"updated_date":"2026-10-26T11:51:55.731654","created_date":"2023-10-04T00:00:00"}