{"id":"https://openalex.org/W4416256166","doi":"https://doi.org/10.48550/arxiv.2509.19943","title":"Interpreting ResNet-based CLIP via Neuron-Attention Decomposition","display_name":"Interpreting ResNet-based CLIP via Neuron-Attention Decomposition","publication_year":2025,"publication_date":"2025-09-24","ids":{"openalex":"https://openalex.org/W4416256166","doi":"https://doi.org/10.48550/arxiv.2509.19943"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2509.19943","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2509.19943","pdf_url":"https://arxiv.org/pdf/2509.19943","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","is_traditional":true,"indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2509.19943","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5120549817","display_name":"Edmund Bu","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I36258959","display_name":"University of California San Diego","ror":"https://ror.org/0168r3w48","country_code":"US","type":"education","lineage":["https://openalex.org/I36258959"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Bu, Edmund","raw_affiliation_strings":["UC San Diego"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"UC San Diego","institution_ids":["https://openalex.org/I36258959"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5051101769","display_name":"Yossi Gandelsman","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I95457486","display_name":"University of California, Berkeley","ror":"https://ror.org/01an7q238","country_code":"US","type":"education","lineage":["https://openalex.org/I95457486"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Gandelsman, Yossi","raw_affiliation_strings":["UC Berkeley"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"UC Berkeley","institution_ids":["https://openalex.org/I95457486"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":2,"corresponding_author_ids":[],"corresponding_institution_ids":[],"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/T12026","display_name":"Explainable Artificial Intelligence (XAI)","score":0.9926999807357788,"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/T12026","display_name":"Explainable Artificial Intelligence (XAI)","score":0.9926999807357788,"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/T10036","display_name":"Advanced Neural Network Applications","score":0.003800000064074993,"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/T11714","display_name":"Multimodal Machine Learning Applications","score":0.0020000000949949026,"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"}}],"keywords":[{"id":"https://openalex.org/keywords/interpretability","display_name":"interpretability","score":0.6169999837875366},{"id":"https://openalex.org/keywords/attention-heads","display_name":"attention heads","score":0.6639999747276306},{"id":"https://openalex.org/keywords/vision-language-models","display_name":"vision-language models","score":0.5910000205039978}],"concepts":[{"id":"https://openalex.org/C41608201","wikidata":"https://www.wikidata.org/wiki/Q980509","display_name":"Embedding","level":2,"score":0.6917999982833862},{"id":"https://openalex.org/C45374587","wikidata":"https://www.wikidata.org/wiki/Q12525525","display_name":"Computation","level":2,"score":0.6827999949455261},{"id":"https://openalex.org/C184898388","wikidata":"https://www.wikidata.org/wiki/Q1435712","display_name":"Pairwise comparison","level":2,"score":0.6653000116348267},{"id":"https://openalex.org/C177264268","wikidata":"https://www.wikidata.org/wiki/Q1514741","display_name":"Set (abstract data type)","level":2,"score":0.6234999895095825},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5821999907493591},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.47600001096725464},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.4603999853134155},{"id":"https://openalex.org/C50644808","wikidata":"https://www.wikidata.org/wiki/Q192776","display_name":"Artificial neural network","level":2,"score":0.4478999972343445},{"id":"https://openalex.org/C153180895","wikidata":"https://www.wikidata.org/wiki/Q7148389","display_name":"Pattern recognition (psychology)","level":2,"score":0.44029998779296875},{"id":"https://openalex.org/C124681953","wikidata":"https://www.wikidata.org/wiki/Q339062","display_name":"Decomposition","level":2,"score":0.3815000057220459},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.3474000096321106},{"id":"https://openalex.org/C110121322","wikidata":"https://www.wikidata.org/wiki/Q865811","display_name":"Distribution (mathematics)","level":2,"score":0.33379998803138733},{"id":"https://openalex.org/C2776359362","wikidata":"https://www.wikidata.org/wiki/Q2145286","display_name":"Representation (politics)","level":3,"score":0.31850001215934753},{"id":"https://openalex.org/C3832189","wikidata":"https://www.wikidata.org/wiki/Q8588916","display_name":"Models of neural computation","level":3,"score":0.29109999537467957},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.2854999899864197},{"id":"https://openalex.org/C2778112365","wikidata":"https://www.wikidata.org/wiki/Q3511065","display_name":"Sequence (biology)","level":2,"score":0.2791000008583069},{"id":"https://openalex.org/C42355184","wikidata":"https://www.wikidata.org/wiki/Q1361088","display_name":"Matrix decomposition","level":3,"score":0.257999986410141}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2509.19943","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2509.19943","pdf_url":"https://arxiv.org/pdf/2509.19943","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.2509.19943","doi":"https://doi.org/10.48550/arxiv.2509.19943","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2509.19943","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:2509.19943","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2509.19943","pdf_url":"https://arxiv.org/pdf/2509.19943","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":[],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4416256166.pdf","grobid_xml":"https://content.openalex.org/works/W4416256166.grobid-xml"},"referenced_works_count":18,"referenced_works":["https://openalex.org/W2194775991","https://openalex.org/W4224302352","https://openalex.org/W4231799185","https://openalex.org/W4303648003","https://openalex.org/W4316135772","https://openalex.org/W4387225729","https://openalex.org/W4387294617","https://openalex.org/W4387560545","https://openalex.org/W4388514540","https://openalex.org/W4389470871","https://openalex.org/W4390873312","https://openalex.org/W4390874324","https://openalex.org/W4396913060","https://openalex.org/W4402716476","https://openalex.org/W4403759200","https://openalex.org/W4407248380","https://openalex.org/W4417097950","https://openalex.org/W4417273915"],"related_works":["https://openalex.org/W4399455353","https://openalex.org/W4403624377","https://openalex.org/W4391248037","https://openalex.org/W4224951374","https://openalex.org/W4387560545","https://openalex.org/W4394710790","https://openalex.org/W4414588339","https://openalex.org/W4415796756","https://openalex.org/W7165363111","https://openalex.org/W7220882618"],"abstract_inverted_index":{"We":[0,40,104],"present":[1],"a":[2,50,76,84],"novel":[3],"technique":[4],"for":[5,108,116,123,159],"interpreting":[6],"the":[7,16,32,80,88,114,128],"neurons":[8,30],"in":[9,53,146],"CLIP-ResNet":[10],"by":[11,49,66],"decomposing":[12],"their":[13,101],"contributions":[14,129],"to":[15,87,133],"output":[17,89],"into":[18],"individual":[19,143],"computation":[20,144],"paths.":[21],"More":[22],"specifically,":[23],"we":[24,61,72,112,126],"analyze":[25],"all":[26],"pairwise":[27],"combinations":[28],"of":[29,36,79,100,130],"and":[31,91,152],"following":[33],"attention":[34],"heads":[35],"CLIP's":[37],"attention-pooling":[38],"layer.":[39],"find":[41,73],"that":[42,74,92,141,153],"these":[43,106],"neuron-head":[44,64,81,94,131],"pairs":[45,82,115,132],"can":[46,156],"be":[47,157],"approximated":[48],"single":[51],"direction":[52],"CLIP-ResNet's":[54],"image-text":[55],"embedding":[56],"space.":[57],"Leveraging":[58],"this":[59],"insight,":[60],"interpret":[62],"each":[63],"pair":[65],"associating":[67],"it":[68],"with":[69],"text.":[70],"Additionally,":[71],"only":[75],"sparse":[77],"set":[78],"have":[83],"significant":[85],"contribution":[86],"value,":[90],"some":[93],"pairs,":[95],"while":[96],"polysemantic,":[97],"represent":[98],"sub-concepts":[99],"corresponding":[102],"neurons.":[103],"use":[105],"observations":[107],"two":[109],"applications.":[110],"First,":[111],"employ":[113],"training-free":[117],"semantic":[118],"segmentation,":[119],"outperforming":[120],"previous":[121],"methods":[122],"CLIP-ResNet.":[124],"Second,":[125],"utilize":[127],"monitor":[134],"dataset":[135],"distribution":[136],"shifts.":[137],"Our":[138],"results":[139],"demonstrate":[140],"examining":[142],"paths":[145],"neural":[147],"networks":[148],"uncovers":[149],"interpretable":[150],"units,":[151],"such":[154],"units":[155],"utilized":[158],"downstream":[160],"tasks.":[161]},"counts_by_year":[],"updated_date":"2026-11-07T11:25:07.637180","created_date":"2025-10-10T00:00:00"}