CLIP-based Few-Shot Multi-Label Classification Methods: A Comparative Study

CLIP-based Few-Sho Mul i-Label Classi ica ion Me hods: A Compa a i e S udy Ya˘ gmu C¸ i˘ gdem Ak as¸ Vicom ech Founda ion Basque Resea ch and Technology Alliance (BRTA) Mikele egi 57, 20009 Donos ia-San Sebas i´ an (Spain) [email p o ec ed]g Jo ge Ga c´ ıa Cas a˜ no Vicom ech Founda ion Basque Resea ch and Technology Alliance (BRTA) Mikele egi 57, 20009 Donos ia-San Sebas i´ an (Spain) [email p o ec ed]g Abs ac —Ca ego izing da kweb image con en is c i ical o iden i ying and a e ing po en ial h ea s. Howe e , his emains a challenge due o he na u e o he da a, which includes mul iple co-exis ing domains and in a-class a ia ions. While many me hods ha e been p oposed o classi y his image con en , mul i-label mul i-class classi ica ion emains unde explo ed. The complexi y o da kweb image y, combined wi h he need o e icien classi ica ion sys ems, demands inno a i e app oaches ha can handle bo h he echnical challenges and he sensi i e na u e o he con en . In his pape , we p esen a compa a i e s udy o ew-sho mul i-label classi ica ion me hods using he mul imodal model CLIP. Ou esea ch add esses he g owing need o obus classi ica ion sys ems ha can e ec i ely ca ego- ize di e se and complex image con en while main aining high accu acy and compu a ional e iciency. We pa icula ly ocus on he challenges o handling mul iple labels simul aneously and he scalabili y o hese sys ems in eal-wo ld applica ions. We analyze and compa e ou di e en app oaches: CLIP+Label Empowe Adap e , CLIP Sigmoid, SIGLIP, and CLIP+ML-Decode . Ou s udy e alua es hese me hods based on hei p ecision, ecall, and abili y o handle inc easing class numbe s e icien ly. Finally, ou esea ch con ibu es o he ield by p o iding de ailed insigh s in o he s eng hs and limi a ions o each me hod. Index Te ms—Mul i-label image classi ica ion, Mul i-class im- age classi ica ion, CLIP, SIGLIP, ML-Decode I. INTRODUCTION The da k web, a specialized subse o he as global in e ne , is accessible only h ough dedica ed web b owse s. I is widely ecognized as a hub o illici ac i i ies due o i s co e cha ac e is ics: anonymi y, which p o ides use s wi h a le el o p i acy no ypically a ailable on he su ace web, and un aceabili y, which makes i ex emely challenging o ack he o igins and des ina ions o da a ans e s. As a esul , he da k web se es as a c i ical sou ce o h ea in elligence, o e ing insigh s in o cybe -a acks, s olen asse s, illegal ade, a ms a icking, con iden ial da a leaks, child exploi a ion, and o he c iminal ac i i ies. Au oma ically analyzing he isual con en o he da k web —such as images and ideos- is essen ial o e icien image ca ego iza ion, enabling he de ec ion and p e en ion o po en ial h ea s. Howe e , da k web image classi ica ion emains an open challenge due o he complex na u e o i s con en . Fig. 1: Mul i-label image examples om Da k web da ase Many images su e om low esolu ion, blu , and small objec sizes, o en con aining objec s ha blend in o he back- g ound due o colo simila i ies. Addi ionally, he coexis ence o mul iple o e lapping domains and in a-class a ia ions — whe e ins ances wi hin he same ca ego y exhibi signi ican isual di e ences — u he complica e classi ica ion asks, as illus a ed in Fig. 1. This complexi y highligh s he impo ance o mul i-label classi ica ion, whe e an image can be assigned mul iple ele- an labels a he han a single ca ego y. A ange o mul i-label classi ica ion echniques ha e been explo ed o ackle his p oblem. Some app oaches decom- pose mul i-label classi ica ion in o sepa a e single-label p ob- lems[1], [2], while o he s e ine loss unc ions o ac i a ion unc ions o enhance classi ica ion accu acy[3], [4]. Addi ion- ally, g aph-based ne wo ks u ilize lea ned label embeddings o loca e key disc imina i e ea u es [5], [6], while ans o me - based a chi ec u es ha e been adap ed o mul i-label classi i- ca ion asks [7]. While hese me hods show p omising esul s, hey ely on la ge, well-balanced da ase s—which a e di icul o cu a e in eal-wo ld applica ions. Expanding a da ase wi h a balanced class dis ibu ion is an exponen ially complex p ocess [8], making la ge-scale mul i-label lea ning in easible in many scena ios. To add ess his, ecen ad ances in ision-language models like CLIP [9] o e new possibili ies. By aligning isual and ex ual embeddings, CLIP enables ze o-sho and ew-sho lea ning in open- ocabula y se ings. Howe e , mos CLIP-979-8-3315-0993-4/25/31.002025IEEE based me hods a e ailo ed o single-label classi ica ion. In his s udy, we add ess he p oblem o ew-sho mul i- label image classi ica ion, ocusing on he da k web as a eal-wo ld, high-s akes applica ion domain. Ou hypo hesis is ha adap ing s a e-o - he-a ew-sho mul i-label classi ica ion echniques o CLIP can imp o e gene aliza ion and pe o - mance in low- esou ce, high- a iabili y en i onmen s. Ou con ibu ions can be summa ized as ollows: •We analyze ou s a e-o - he-a ew-sho mul i-label classi ica ion echniques and p o ide a po able pipeline applicable o any eal-wo ld da ase equi ing ew-sho mul i-label classi ica ion. •We e alua e hese app oaches by implemen ing CLIP- based me hodologies o da k web mul i-label image classi ica ion. •We compa e ou p e ious app oach wi h newe CLIP- based mul i-label me hodologies, p esen ing a anspa en compa a i e s udy which con ibu es o he ield by p o- iding de ailed insigh s in o he s eng hs and limi a ions o each me hod. II. RELATED WORK A a ie y o me hods ha e been p oposed o sol e he mul i- label classi ica ion p oblem, which can be b oadly ca ego ized in o (i) p oblem ans o ma ion me hods and (ii) algo i hm adap a ion echniques. Ea ly esea ch ocused on decomposing mul i-label classi i- ca ion in o independen bina y classi ica ion asks[1]. How- e e , his app oach ails o cap u e in e -label co ela ions and equi es aining sepa a e classi ie s o each ca e- go y—in oducing a signi ican compu a ional bu den. To ad- d ess his, some s udies p opose classi ie chaining[10], whe e each classi ie ’s ou pu se es as an inpu ea u e o sub- sequen classi ie s. While his s a egy imp o es in e -label dependency modeling, i s ill su e s om high compu a ional cos s due o he g owing numbe o classi ie s equi ed. Ano he ans o ma ion-based app oach, known as Label Powe se (LP) [2], con e s mul i-label p oblems in o mul i- class p oblems by ea ing each unique label combina ion as a dis inc class. Howe e , LP s uggles wi h he exponen ial g ow h o label combina ions, making i imp ac ical o la ge da ase s. Se e al me hods aim o imp o e mul i-label classi ica ion by op imizing he loss unc ion[3] o ac i a ion unc ion[4] o mi iga e class imbalance issues. In compu e ision, mul iple app oaches ha e been de el- oped o mul i-label image classi ica ion. Hypo heses-CNN- Pooling (HCP) [11] gene a es a la ge numbe o p oposals h ough objec de ec ion echniques, ea ing each p oposal as a single-label classi ica ion p oblem. Beyond CNN-based me hods, G aph Con olu ional Ne - wo ks (GCNs) ha e demons a ed high e icacy ac oss a i- ous ision asks, including mul i-label classi ica ion [5], [6], [12]. GCN-based me hods build classi ie s by modeling label ela ionships wi hin g aph ne wo ks. O he app oaches employ weakly supe ised lea ning o mul i-label classi ica ion [13], le e aging knowledge dis illa- ion echniques om objec de ec ion models. Mo e ecen ly, ans o me s[14]—o iginally de eloped o modeling long- ange dependencies in na u al language p o- cessing (NLP)[15]–[17]—ha e demons a ed s ong pe o - mance ac oss compu e ision asks, including image classi i- ca ion[7], [9] and objec de ec ion[18]. Se e al ans o me -based mul i-label classi ica ion ech- niques ha e eme ged: Mul i-label T ans o me (MLT) [19], which models pixel- wise a en ion o enhance ea u e ex ac ion. T ans o me - based que y lea ning [20], which uses decode que ies o p edic label exis ence. G aph-enhanced T ans o me s [21], which combine GCNs wi h a en ion mechanisms. Me ic Lea ning T ans o me s [22], which in eg a e me ic lea ning o assess label simila i ies. T ans o me -CNN Hyb ids [23], which gene a e a en ion maps pe label, cap u ing in a-class dependencies h ough con olu ional laye s. Despi e hese ad ancemen s, exis ing mul i-label me hods s ill equi e la ge da ase s o achie e high accu acy. Howe e , da a sca ci y is a p e alen challenge in eal-wo ld applica- ions, whe e collec ing and labeling ex ensi e da ase s is bo h esou ce-in ensi e and ime-consuming. La ge-scale open- ocabula y models ained on as da ase s, such as CLIP [9], o e a p omising al e na i e by le e aging ze o-sho single-label classi ica ion and ew-sho mul i-label classi ica ion. By adap ing hese models, we aim o explo e hei po en ial o enhancing mul i-label classi i- ca ion in da k web image y, whe e adi ional me hods ace signi ican cons ain s. III. CLIP-BASED MULTI-LABEL CLASSIFICATION METHODOLOGIES The e a e ou ecen mul i-label classi ica ion me hods sui - able o CLIP mul i-modal model adap a ion: (i) CLIP+Label Empowe (ou p e ious model), (ii) CLIP Sigmoid (iii) SIGLIP (i ) CLIP + ML-Decode A. CLIP+Label Empowe Label Empowe [2] is one o he app oaches ha con e he mul i-label classi ica ion p oblem in o a single-label classi i- ca ion, by mul i-labels o bina y ec o s by ho -encoding hem and assigning a unique label o each di e en bina y ec o . As demons a ed in [24], i has he highes pe o mance among he o he app oaches o sol ing he mul i-label classi ica ion ask by ans o ming he p oblem in o a single-label one. Fig. 2 shows he a chi ec u e o ou p e iously p oposed me hod CLIP+Label Empowe Adap e . Al hough ou p e iously p oposed model has he highes ecall, as shown in Table I, among e en hese ou s a e- o - he-a CLIP-adap ed mul i-label classi ica ion app oaches, no only be ween he adi ional ones, i p ese es a isk o accu acy d op wi hin a la ge numbe o classes due o i s exponen ially g owing ho -encoded labels. This possible isk is especially impo an o such a da ase like da kweb, which Fig. 2: CLIP based Mul i-Label Me hodologies is expec ed o g ow as in he sense o numbe o classes, since new ype o images and i les a e expec ed o occu . •Wo s Case: O(N!) (All class labels exis in all possible pe mu a ions in he da ase ) •Bes Case: O(N)(The da ase con ains only single labels) B. CLIP Sigmoid CLIP has he capabili y o classi y an ex emely huge num- be o di e en classes, hanks o i s open- ocabula y na u e. Ye we migh need o ine- une his s a e-o - he-a model o ou eal-wo ld da ase s like da kweb da ase . In a single-label pipeline, he p ope way o ine- uning such a model is o add a linea laye ollowing he image and ex embeddings, ha ing ou pu nodes as much as class numbe s in he da ase . In a single-label, mul i-class classi ica ion scena io, so max is commonly used o dis ibu e p obabili ies among di e en classes, ensu ing ha he sum o all class p obabili ies equals 1. P(yi|x) = ezi PN j=1 ezj (1) whe e zi ep esen s he logi s o class iand Nis he o al numbe o classes. Howe e , in mul i-label classi ica ion, whe e mul iple labels can be p esen simul aneously, so max is subop imal as i o ces mu ual exclusi i y among classes. Ins ead, sigmoid ac i a ion is mo e app op ia e, as i independen ly p edic s each label’s p obabili y: P(yi|x) = 1 1 + e−zi(2) This allows he model o assign a p obabili y close o 1.0 o each ele an label in an image, ensu ing ha all p esen classes a e p edic ed wi hou compe i ion om o he labels. Thus, ine- uning CLIP wi h a linea classi ica ion head using Sigmoid ac i a ion is ano he app oach. Al hough his me hod does no ha e any isk o exponen ially g owing wi h he numbe o class numbe s, ou expe imen al esul s show i s poo accu acy in compa ison wi h ou p e iously p oposed me hod. C. SIGLIP While SIGLIP [25] is a s a e-o - he-a , mul i-modal model, ha ing a e y simila a chi ec u e o CLIP, i s main di e ence is using Sigmoid, ins ead o So max while p e- aining. This ac p o ides us wi h he possibili y o see i he poo pe o - mance o CLIP + Sigmoid ine- uning me hodology occu ed due o using a di e en ac i a ion unc ion in he ine- uning han p e- aining. Fig. 2 shows he e y simila a chi ec u e o SIGLIP. D. CLIP+ML-Decode ML-Decode [26] is a s a e-o - he-a decode implemen- a ion aiming o o e come he exponen ially g owing inpu size in classi ica ion pipelines buil wi h de aul ans o me decode s (Fig. 3) due o he sel -a en ion laye . The me hod p oposes o emo e he sel -a en ion laye and claims i s a ec less on he classi ica ion accu acy. I is also claimed o use ”g oup que ies”, ins ead o ixed que y embeddings pe class as in adi ional ans o me -decode s. Being he numbe o g oups a new hype pa ame e in his a chi ec u e e e s o he amoun o inpu que y embeddings he decode will ha e, independen ly om he numbe o classes in he da ase . Being K is he numbe o g oup que ies, he decode lea ns o c ea e K que ies e e encing N numbe o classes in a da ase , ins ead o ha ing N ixed que ies o N classes as in de aul ans o me decode s. CLIP+ML-Decode app oach uses bo h classi ica ion loss coming om he inal p edic ion logi s and an alignmen loss coming om he simila i y o ex and image embeddings o ob ain a inal loss. The ixed ex embeddings a e used only o his pu pose, whe eas non- ixed g oup que ies a e lea nable embeddings and hey a e upda ed du ing he aining. Fig. 3: T adi ional s ML Decode s The e o e his app oach enhances scalabili y, wi h wo key op imiza ions: (1) Sel -a en ion emo al, which educes he compu a ional complexi y om O(N2) o O(N), and (2) G oup decoding, which u he op imizes in e ence by shi ing he complexi y om O(N) o O(K), whe e KK ep esen s he numbe o meaning ul g oups ins ead o p ocessing all classes independen ly. Sel -a en ion emo al: O(N2)→O(N) G oup decoding: O(N)→O(K) Fig. 2 shows he a chi ec u e o CLIP + ML-Decode . IV. EXPERIMENTAL STUDY A. Da k web da ase The Da k web da ase , sou ced om CFLW’s Da k Web Moni o 1, includes images collec ed om da k web domains abou a ious c ime ca ego ies like inancial c ime and o gani- za ions, d ugs and na co ics, weapons as well as hei endo s as indi idual classes. Simila o many eal-wo ld da ase s, he da k web da ase con ains images wi h mul iple labels, meaning an image can belong o mo e han one class. I also includes images wi h a single label. Fig. 1 shows some image samples along wi h hei g ound u h labels om a ious classes. 1h ps://c lw.com/dwm/ Fig. 4: Da a dis ibu ion o he da k web da ase o bo h ain and es se s. The da ase con ains 46 classes om a ious ca ego ies like D ugs, Na co ics, Weapons, and Financial O ganiza ions which do no ha e a s ong co ela ion be ween hem bu i includes subca ego ies ha ing a s ong ela ionship and he possibili y o being exis ing in he same image. Being he cu en da kweb da ase is an expe imen al one, many mo e classes a e expec ed o join and hus, he scalabili y pe o mance is impo an han a egula Fig. 4 shows he da a dis ibu ion o he da k web da ase o bo h ain and es se s. The da kweb da ase p esen s an imbalanced da a p oblem. Some dominan classes ha e many samples, whe eas some o he classes su e om a lack o da a. Some classes like d ugs and na co ics, coming along wi h any ype o d ug ype o any indi idual endo ha has o be ca ego ized, become a e y dominan class by collec ing only a ew samples o i s subg oups. Fig. 1 shows 3 image samples om he da k web da ase , 1 belonging o a inancial o ganiza ion ca ego y, 2 belonging o he d ugs and na co ics ca ego y con aining pic u es o weed and he endo names. While endo names may a y, each image ha ing a endo name in o ma ion also con ains weed class, doubling he amoun o ”weed” class in compa ison o class names e e ing o he indi idual endo names. Which is obus example showing he eason o s ong class imbalance p oblem in da kweb da ase . Las ly, ano he challenge is ha some classes, especially he endo names like ”deep shop”, o ” anda al”, ha e ex ual in o ma ion a he han isual ea u es which b ings he need o conside ing he combina ion o ision and language da a. B. Implemen a ion De ails The ine- uning s a egy a ies o each o he me hods we examine: CLIP + Label Empowe was ine uned by eezing he CLIP p e- ained model, and upda ing he weigh s o only o be ween 5 o 10 epochs, using Adam op imize wi h de aul lea ning a e 1×10−3. So max inal ac i a ion unc ion and C oss En opy Loss. CLIP Sigmoid model was ine uned by eezing he CLIP p e- ained model, and upda ing he weigh s o only he ad- di ional linea classi ica ion head, o 50 epochs un il con e - gence, using he same op imize and lea ning a e as p e ious app oach, wi h a Bina y C oss En opy Loss o e he logi s, as a de aul app oach o sigmoid classi ica ion. SIGLIP model was ine uned wi hou eezing any pa o he a chi ec u e, since ou implemen a ion any addi ional pa and he na u e o he a chi ec u e is al eady compa ible o mul i-label classi ica ion. I is ine uned 80 epochs un il con e gence, using Adam op imize wi h a lea ning a e o 5×10−5. CLIP+ML-Decode me hod was ine- uned by eezing he CLIP model and only ocusing on upda ing he weigh s o Decode . A g id sea ch o e all he possible hype -pa ame e s o he decode block was made and he bes esul s was ob ained wi h: mul i-head a en ion head amoun 4, d opou in eed o wa d ne wo k 0.5, numbe o g oups K 8, numbe o laye s (decode block amoun ) 2. Simila ly o he p e ious me hods, Adam op imize is used, wi h a lea ning a e o 1×10−4. C. E alua ion Me ics We e alua e model pe o mance using s anda d classi ica- ion me ics: p ecision, ecall, and 1-sco e. We u he mo e analyze he scalabili y o he model. •P ecision: Measu es he p opo ion o co ec ly p edic ed labels among all p edic ed labels. 3 •Recall: Measu es he p opo ion o co ec ly p edic ed labels among all ue labels.4 •F1-Sco e: The ha monic mean o p ecision and ecall. 5 •Scalabili y: Assesses how well he model adap s o inc easing class sizes. P ecision =T P T P +F P (3) Recall =T P T P +F N (4) F1sco e =2×P ecision ×Recall P ecision +Recall (5) The ue posi i es, alse posi i es, alse nega i es a e cal- cula ed simila ly o single-label classi ica ion. Fig. 5 shows an image sample ha ing wo g ound u h labels: VISA and Bankno es. In case he p edic ed classes a e ”VISA and Paypal” class, a e ex ac ing he single labels as ”VISA” and ”Paypal”, his p edic ion would con ibu e as a ue posi i e o ”VISA” class, alse posi i e o ”Paypal” class and a alse nega i e o ”Bankno es” classes. Fig. 5: An example mul i-labeled image wi h 2 classes. V. RESULTS AND DISCUSSION Table I compa es he pe o mance o he ou me hods in e ms o p ecision, ecall, 1-sco e and scalabili y and summa izes ou indings. CLIP+Label Empwoe me hod has a low scalabili y since i has O(N!) as wo s case scena io, CLIP Sigmoid and SIGLIP has a mode a e scalabili y since hey p o ide an imp o emen , bu nei he b ing any g ow h, meaning he wo s and bes case scena io is he same and O(N). While CLIP+ML-Decode imp o es he scalabili y om O(N) o O(K), being N is he numbe o classes and K is he numbe o g oups de ined by he end-use . Table II compa es he wo models gi ing he bes esul s o da k web da ase , on he MS-COCO mul i-label da ase , causing he CLIP+LE me hod o ha e a huge d op o accu acy, due o expanding 80 base classes o 234,581 ho -encoded classes. These esul s show he a o emen ioned po en ial isk o CLIP+LE me hod on he da ase s ha ing high ela ion be ween he classes, causing he exis ence o a ious pe mu- a ions o mul i-label ec o s among he da ase . E en hough he esul s migh be imp essi e o such amoun o classes, i is isible ha he Label Empowe me hod has a huge impac on CLIP mul i-modal model’s classi ica ion capaci y by exploding he class numbe s unnecessa ily. TABLE I COMPARISON OF FEW-SHOT MULTI-LABEL CLASSIFICATION METHODS. LE: LABEL EMPOWER, MLD: ML-DECODER Me hod P ecision Recall F1 sco e Scalabili y CLIP+LE 0.944 0.931 0.937 Low CLIP Sigmoid 0.493 0.276 0.351 Mode a e SIGLIP 0.880 0.735 0.801 Mode a e CLIP+MLD 0.958 0.916 0.936 High Ou esul s show ha he CLIP+Label Empowe Adap e achie es he bes ecall, making i s ill a s ong candida e o ecall-sensi i e applica ions buil o da ase s no ha ing huge amoun o classes, due o i s low scalabili y and he po en ial isk o p o iding poo e esul s wi h a huge numbe o classes. On he o he hand, he esul s show he medium- le el scalabili y models (CLIP+Sigmoid, SIGLIP), ha a e no imp o ing no wo sen he a chi ec u e acco ding o he numbe o classes, ha e poo accu acy in compa ison wi h CLIP + adap e solu ions. Howe e , CLIP+ML-Decode pe o ms well in bo h p ecision and ecall, while also add essing he scalabili y issue e ec i ely and i should be a de ini e choice o da ase ha ing huge numbe o classes, whe e o no ha e alse posi i e p edic ions is mo e impo an han no missing any ue posi i e p edic ion, ega ding i ’s p ecision bea ing he CLIP+Label Empowe me hod while p o iding lowe ecall. TABLE II COMPARISON OF MULTI-LABEL MS-COCO MAP SCORE. LE: LABEL EMPOWER AND MLD: MULTI-LABEL DECODER Me hod P ecision Recall CLIP+LE 0.592 0.555 CLIP+MLD 0.839 0.809 VI. CONCLUSION AND FUTURE WORKS In his s udy, we analyzed ou ew-sho mul i-label classi- ica ion me hods based on CLIP. Ou esul s demons a e ha CLIP+Label Empowe Adap e excels in he ecall, whe eas CLIP+ML-Decode p o ides a mo e scalable solu ion by mi iga ing he exponen ial g ow h p oblem in inpu ea u es, p o iding also a obus pe o mance on accu acy me ics. Fu u e wo k will explo e he e icien deploymen pipeline o CLIP+ML-Decode app oach o eal-wo ld mul i-label classi ica ion asks. ACKNOWLEDGEMENTS The wo k desc ibed in his pape is pe - o med in he H2020 p ojec STARLIGHT (”Sus ainable Au onomy and Resilience o LEAs using AI agains High P io i y Th ea s”). This p ojec has ecei ed unding om he Eu opean Union’s Ho izon 2020 esea ch and inno a ion p og am unde g an ag eemen No 101021797. REFERENCES [1] M.-L. Zhang, Y.-K. Li, X.-Y. Liu, and X. Geng, “Bina y ele ance o mul i-label lea ning: An o e iew,” F on ie s o Compu e Science, ol. 12, no. 2, pp. 191–202, Ma . 2018, ISSN: 2095-2236. DOI: 10. 1007/s11704-017-7031-7. [Online]. A ailable: h p://dx.doi.o g/10. 1007/s11704-017-7031-7. [2] G. Tsoumakas, A. Dimou, E. Spy omi os-Xiou is, V. Meza is, I. Kompa sia is, and I. Vlaha as, “Co ela ion-based p uning o s acked bina y ele ance models o mul i-label lea ning,” Jan. 2009, pp. 101– 116. [3] E. Ben-Ba uch, T. Ridnik, N. Zami , e al.,Asymme ic loss o mul i- label classi ica ion, 2021. a Xi : 2009.14119 [cs.CV]. [4] A. F. T. Ma ins and R. F. As udillo, F om so max o spa semax: A spa se model o a en ion and mul i-label classi ica ion, 2016. a Xi : 1602.02068 [cs.CL]. [5] Y. Wang, D. He, F. Li, e al.,Mul i-label classi ica ion wi h label g aph supe imposing, 2019. a Xi : 1911.09243 [cs.CV]. [6] Z.-M. Chen, X.-S. Wei, P. Wang, and Y. Guo, Mul i-label image ecogni ion wi h g aph con olu ional ne wo ks, 2019. a Xi : 1904. 03582 [cs.CV]. [7] H. Tou on, M. Co d, M. Douze, F. Massa, A. Sablay olles, and H. J´ egou, T aining da a-e icien image ans o me s dis illa ion h ough a en ion, 2021. a Xi : 2012.12877 [cs.CV]. [8] W. Zhang, C. Liu, L. Zeng, B. Ooi, S. Tang, and Y. Zhuang, “Lea ning in impe ec en i onmen : Mul i-label classi ica ion wi h long- ailed dis ibu ion and pa ial labels,” in P oceedings o he IEEE/CVF In e na ional Con e ence on Compu e Vision (ICCV), Oc . 2023, pp. 1423–1432. [9] A. Rad o d, J. W. Kim, C. Hallacy, e al.,Lea ning ans e able isual models om na u al language supe ision, 2021. a Xi : 2103.00020 [cs.CV]. [10] J. Read, B. P ah inge , G. Holmes, and E. F ank, “Classi ie chains: A e iew and pe spec i es,” Jou nal o A i icial In elligence Resea ch, ol. 70, pp. 683–718, Feb. 2021, ISSN: 1076-9757. DOI: 10.1613/jai . 1.12376. [Online]. A ailable: h p://dx.doi.o g/10.1613/jai .1.12376. [11] Y. Wei, W. Xia, M. Lin, e al., “Hcp: A lexible cnn amewo k o mul i-label image classi ica ion,” IEEE T ansac ions on Pa e n Analysis and Machine In elligence, ol. 38, no. 9, pp. 1901–1907, Sep. 2016, ISSN: 1939-3539. DOI: 10 . 1109 / pami . 2015 . 2491929. [Online]. A ailable: h p://dx.doi.o g/10.1109/TPAMI.2015.2491929. [12] T. Chen, M. Xu, X. Hui, H. Wu, and L. Lin, Lea ning seman ic- speci ic g aph ep esen a ion o mul i-label image ecogni ion, 2019. a Xi : 1908.07325 [cs.CV]. [13] Y. Liu, L. Sheng, J. Shao, J. Yan, S. Xiang, and C. Pan, “Mul i- label image classi ica ion ia knowledge dis illa ion om weakly- supe ised de ec ion,” in P oceedings o he 26 h ACM in e na ional con e ence on Mul imedia, ACM, Oc . 2018. DOI: 10.1145/3240508. 3240567. [Online]. A ailable: h p://dx.doi.o g/10.1145/3240508. 3240567. [14] A. Vaswani, N. Shazee , N. Pa ma , e al.,A en ion is all you need, 2023. a Xi : 1706.03762 [cs.CL]. [15] J. De lin, M.-W. Chang, K. Lee, and K. Tou ano a, Be : P e- aining o deep bidi ec ional ans o me s o language unde s anding, 2019. a Xi : 1810.04805 [cs.CL]. [16] A. Rad o d, J. Wu, R. Child, D. Luan, D. Amodei, and I. Su ske e , “Language models a e unsupe ised mul i ask lea ne s,” 2019. [On- line]. A ailable: h ps://api.seman icschola .o g/Co pusID:160025533. [17] T. B. B own, B. Mann, N. Ryde , e al.,Language models a e ew- sho lea ne s, 2020. a Xi : 2005.14165 [cs.CL]. [18] X. Zhu, W. Su, L. Lu, B. Li, X. Wang, and J. Dai, De o mable de : De o mable ans o me s o end- o-end objec de ec ion, 2021. a Xi : 2010.04159 [cs.CV]. [19] X. Cheng, H. Lin, X. Wu, e al.,Ml : Mul i-label classi ica ion wi h ans o me , 2021. a Xi : 2106.06195 [cs.CV]. [20] S. Liu, L. Zhang, X. Yang, H. Su, and J. Zhu, Que y2label: A simple ans o me way o mul i-label classi ica ion, 2021. a Xi : 2107.10834 [cs.CV]. [21] J. Ye, J. He, X. Peng, W. Wu, and Y. Qiao, A en ion-d i en dynamic g aph con olu ional ne wo k o mul i-label image ecogni ion, 2020. a Xi : 2012.02994 [cs.CV]. [22] K. P oko ie and V. So aso , Combining me ic lea ning and a en- ion heads o accu a e and e icien mul ilabel image classi ica ion, 2022. a Xi : 2209.06585 [cs.CV]. [23] F. Zhu, H. Li, W. Ouyang, N. Yu, and X. Wang, Lea ning spa ial egula iza ion wi h image-le el supe isions o mul i-label image classi ica ion, 2017. a Xi : 1702.05891 [cs.CV]. [24] Y. C¸ . Ak as¸ and J. G. Cas a˜ no, “Few-sho mul i-label mul i-class classi ica ion o da k web image ca ego iza ion,” in 2024 12 h In e na ional Symposium on Digi al Fo ensics and Secu i y (ISDFS), 2024, pp. 1–6. DOI: 10.1109/ISDFS60797.2024.10527297. [25] X. Zhai, B. Mus a a, A. Kolesniko , and L. Beye , Sigmoid loss o language image p e- aining, 2023. a Xi : 2303 .15343 [cs.CV]. [Online]. A ailable: h ps://a xi .o g/abs/2303.15343. [26] T. Ridnik, G. Sha i , A. Ben-Cohen, E. Ben-Ba uch, and A. Noy, Ml-decode : Scalable and e sa ile classi ica ion head, 2021. a Xi : 2111.12933 [cs.CV]. [Online]. A ailable: h ps://a xi .o g/abs/ 2111.12933.