{"id":"https://openalex.org/W7052201220","doi":"https://doi.org/10.48550/arxiv.2307.00331","title":"Quantization Variation: A New Perspective on Training Transformers with Low-Bit Precision","display_name":"Quantization Variation: A New Perspective on Training Transformers with Low-Bit Precision","publication_year":2023,"publication_date":"2023-07-01","ids":{"openalex":"https://openalex.org/W7052201220","doi":"https://doi.org/10.48550/arxiv.2307.00331"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2307.00331","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2307.00331","pdf_url":"https://arxiv.org/pdf/2307.00331","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":"cc-by-nc-sa","license_id":"https://openalex.org/licenses/cc-by-nc-sa","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/2307.00331","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5049883737","display_name":"Xijie Huang","orcid":"https://orcid.org/0000-0001-5508-5433","observed_orcids":["https://orcid.org/0000-0001-5508-5433"]},"institutions":[{"id":"https://openalex.org/I200769079","display_name":"Hong Kong University of Science and Technology","ror":"https://ror.org/00q4vv597","country_code":"HK","type":"education","lineage":["https://openalex.org/I200769079"]}],"countries":["HK"],"is_corresponding":false,"raw_author_name":"Huang, Xijie","raw_affiliation_strings":["Hong Kong University of Science and Technology ,"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Hong Kong University of Science and Technology ,","institution_ids":["https://openalex.org/I200769079"]}]},{"author_position":"middle","author":{"id":null,"display_name":"Shen, Zhiqiang","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I4210113480","display_name":"Mohamed bin Zayed University of Artificial Intelligence","ror":"https://ror.org/0258gkt32","country_code":"AE","type":"education","lineage":["https://openalex.org/I4210113480"]}],"countries":["AE"],"is_corresponding":false,"raw_author_name":"Shen, Zhiqiang","raw_affiliation_strings":["Mohamed bin Zayed University of Artificial Intelligence"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Mohamed bin Zayed University of Artificial Intelligence","institution_ids":["https://openalex.org/I4210113480"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5043072381","display_name":"Pingcheng Dong","orcid":"https://orcid.org/0000-0003-3975-286X","observed_orcids":["https://orcid.org/0000-0003-3975-286X"]},"institutions":[{"id":"https://openalex.org/I200769079","display_name":"Hong Kong University of Science and Technology","ror":"https://ror.org/00q4vv597","country_code":"HK","type":"education","lineage":["https://openalex.org/I200769079"]}],"countries":["HK"],"is_corresponding":false,"raw_author_name":"Dong, Pingcheng","raw_affiliation_strings":["Hong Kong University of Science and Technology ,"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Hong Kong University of Science and Technology ,","institution_ids":["https://openalex.org/I200769079"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5077687075","display_name":"Kwang-Ting Tim Cheng","orcid":"https://orcid.org/0000-0002-3885-4912","observed_orcids":["https://orcid.org/0000-0002-3885-4912"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Cheng, Kwang-Ting","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]}],"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":16,"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":true,"primary_topic":{"id":"https://openalex.org/T10036","display_name":"Advanced Neural Network Applications","score":0.9984999895095825,"subfield":{"id":"https://openalex.org/subfields/1707","display_name":"Computer Vision and Pattern Recognition"},"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/T10036","display_name":"Advanced Neural Network Applications","score":0.9984999895095825,"subfield":{"id":"https://openalex.org/subfields/1707","display_name":"Computer Vision and Pattern Recognition"},"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":0.0003000000142492354,"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":9.999999747378752e-05,"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/low-bit-quantization","display_name":"low-bit quantization","score":0.8320000171661377},{"id":"https://openalex.org/keywords/quantization-aware-training","display_name":"quantization-aware training","score":0.9319999814033508},{"id":"https://openalex.org/keywords/transformer","display_name":"transformer models","score":0.9179999828338623},{"id":"https://openalex.org/keywords/model-compression","display_name":"model compression","score":0.7940000295639038},{"id":"https://openalex.org/keywords/vision-transformer","display_name":"vision transformer","score":0.5720000267028809},{"id":"https://openalex.org/keywords/parameter-fluctuations","display_name":"parameter fluctuations","score":0.5120000243186951}],"concepts":[{"id":"https://openalex.org/C28855332","wikidata":"https://www.wikidata.org/wiki/Q198099","display_name":"Quantization (signal processing)","level":2,"score":0.6721000075340271},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6495000123977661},{"id":"https://openalex.org/C66322947","wikidata":"https://www.wikidata.org/wiki/Q11658","display_name":"Transformer","level":3,"score":0.5940999984741211},{"id":"https://openalex.org/C45374587","wikidata":"https://www.wikidata.org/wiki/Q12525525","display_name":"Computation","level":2,"score":0.583899974822998},{"id":"https://openalex.org/C79337645","wikidata":"https://www.wikidata.org/wiki/Q779824","display_name":"Outlier","level":2,"score":0.423799991607666},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.3856000006198883},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.3610000014305115},{"id":"https://openalex.org/C48372109","wikidata":"https://www.wikidata.org/wiki/Q3913","display_name":"Binary number","level":2,"score":0.3409999907016754},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.34040001034736633},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.2858999967575073},{"id":"https://openalex.org/C51632099","wikidata":"https://www.wikidata.org/wiki/Q3985153","display_name":"Training set","level":2,"score":0.28200000524520874},{"id":"https://openalex.org/C113775141","wikidata":"https://www.wikidata.org/wiki/Q428691","display_name":"Computer engineering","level":1,"score":0.2775000035762787},{"id":"https://openalex.org/C199833920","wikidata":"https://www.wikidata.org/wiki/Q612536","display_name":"Vector quantization","level":2,"score":0.2590000033378601}],"mesh":[],"locations_count":3,"locations":[{"id":"pmh:oai:arXiv.org:2307.00331","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2307.00331","pdf_url":"https://arxiv.org/pdf/2307.00331","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":"cc-by-nc-sa","license_id":"https://openalex.org/licenses/cc-by-nc-sa","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"pmh:oai:repository.hkust.edu.hk:1783.1-148583","doi":null,"is_oa":false,"landing_page_url":"http://www.scopus.com/record/display.url?eid=2-s2.0-105000719675&origin=inward","pdf_url":null,"source":{"id":"https://openalex.org/S4306401796","display_name":"Rare & Special e-Zone (The Hong Kong University of Science and Technology)","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I200769079","host_organization_name":"Hong Kong University of Science and Technology","host_organization_lineage":["https://openalex.org/I200769079"],"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":"Article"},{"id":"doi:10.48550/arxiv.2307.00331","doi":"https://doi.org/10.48550/arxiv.2307.00331","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2307.00331","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:2307.00331","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2307.00331","pdf_url":"https://arxiv.org/pdf/2307.00331","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":"cc-by-nc-sa","license_id":"https://openalex.org/licenses/cc-by-nc-sa","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"sustainable_development_goals":[],"study_designs":[],"awards":[{"id":"https://openalex.org/G2381494637","display_name":null,"funder_award_id":"N_HKUST627/20","funder_id":"https://openalex.org/F4320323537","funder_display_name":"Hong Kong University of Science and Technology"},{"id":"https://openalex.org/G481671389","display_name":null,"funder_award_id":"N_HKUST627/20","funder_id":"https://openalex.org/F4320321001","funder_display_name":"National Natural Science Foundation of China"}],"funders":[{"id":"https://openalex.org/F4320321001","display_name":"National Natural Science Foundation of China","ror":"https://ror.org/01h0zpd94"},{"id":"https://openalex.org/F4320323537","display_name":"Hong Kong University of Science and Technology","ror":"https://ror.org/00q4vv597"}],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7052201220.pdf","grobid_xml":"https://content.openalex.org/works/W7052201220.grobid-xml"},"referenced_works_count":24,"referenced_works":["https://openalex.org/W1821462560","https://openalex.org/W2242818861","https://openalex.org/W2612445135","https://openalex.org/W2896457183","https://openalex.org/W2965373594","https://openalex.org/W3037932933","https://openalex.org/W3170437634","https://openalex.org/W3212960520","https://openalex.org/W4224308101","https://openalex.org/W4225857786","https://openalex.org/W4226174177","https://openalex.org/W4285601701","https://openalex.org/W4312480718","https://openalex.org/W4313069943","https://openalex.org/W4327810158","https://openalex.org/W4378770729","https://openalex.org/W4379260375","https://openalex.org/W4384918448","https://openalex.org/W4385819842","https://openalex.org/W4389524393","https://openalex.org/W4390873361","https://openalex.org/W4390874575","https://openalex.org/W4391157567","https://openalex.org/W7133229067"],"related_works":["https://openalex.org/W3202028501","https://openalex.org/W4392737200","https://openalex.org/W7162441734","https://openalex.org/W7133267101","https://openalex.org/W3216998657","https://openalex.org/W4386158858","https://openalex.org/W4386566804","https://openalex.org/W7034215471","https://openalex.org/W7133946636","https://openalex.org/W4416264540"],"abstract_inverted_index":{"Despite":[0],"the":[1,12,20,27,38,56,61,75,127,134,138,143,171,175,188],"outstanding":[2],"performance":[3,176],"of":[4,59,77,122,177],"transformers":[5,123,178],"in":[6,37,100,108,116],"both":[7,157],"language":[8,160],"and":[9,15,30,46,68,111,114,131,150,159,165,173,182,191,197,207,210],"vision":[10,158],"tasks,":[11],"expanding":[13],"computation":[14,29],"model":[16,43],"size":[17],"have":[18],"increased":[19],"demand":[21],"for":[22,156],"efficient":[23],"deployment.":[24],"To":[25],"address":[26],"heavy":[28],"parameter":[31,118],"drawbacks,":[32],"quantization":[33,63,105,154],"is":[34],"frequently":[35],"studied":[36],"community":[39],"as":[40],"a":[41,152,195],"representative":[42],"compression":[44],"technique":[45],"has":[47],"seen":[48],"extensive":[49],"use":[50],"on":[51,82,93,205],"ConvNets.":[52,91],"However,":[53],"due":[54],"to":[55,126,141],"unique":[57,84],"properties":[58],"transformers,":[60],"low-bit":[62,79,147],"applications":[64],"are":[65,212],"still":[66],"limited":[67],"underexplored.":[69],"In":[70],"this":[71],"paper,":[72],"we":[73,97],"identify":[74],"difficulty":[76],"transformer":[78,148],"quantization-aware":[80,128],"training":[81,129],"its":[83],"variation":[85,99,172],"behaviors,":[86],"which":[87],"significantly":[88],"differ":[89],"from":[90],"Based":[92],"comprehensive":[94],"quantitative":[95],"analysis,":[96],"observe":[98],"three":[101],"hierarchies:":[102],"various":[103,180],"module":[104],"sensitivities,":[106],"outliers":[107],"static":[109],"weight":[110],"activation":[112],"distribution,":[113],"oscillation":[115],"dynamic":[117],"fluctuations.":[119],"These":[120],"variations":[121],"bring":[124],"instability":[125],"(QAT)":[130],"negatively":[132],"influence":[133,145],"performance.":[135],"We":[136,162],"explore":[137],"best":[139],"practices":[140],"alleviate":[142,170],"variation's":[144],"during":[146],"QAT":[149],"propose":[151],"variation-aware":[153],"scheme":[155,168],"transformers.":[161],"extensively":[163],"verify":[164],"show":[166],"our":[167],"can":[169],"improve":[174],"across":[179],"models":[181,211],"tasks.":[183],"Our":[184],"solution":[185],"substantially":[186],"improves":[187],"2-bit":[189],"Swin-T":[190],"binary":[192],"BERT-base,":[193],"achieving":[194],"3.35%":[196],"1.4%":[198],"accuracy":[199],"improvement":[200],"over":[201],"previous":[202],"state-of-the-art":[203],"methods":[204],"ImageNet-1K":[206],"GLUE.":[208],"Codes":[209],"available":[213],"at":[214],"https://github.com/HuangOwen/Quantization-Variation.":[215]},"counts_by_year":[{"year":2026,"cited_by_count":2},{"year":2025,"cited_by_count":6},{"year":2024,"cited_by_count":8}],"updated_date":"2026-11-08T12:12:28.789561","created_date":"2025-10-10T00:00:00"}