Kwenzekeni
Abacwaningi bethule i-VGA-BenchV2, ibhentshimakhi enwetshiwe kanye nohlaka lokuthuthukisa lokuhlola ividiyo ekhiqizwe yi-AI. Uhlelo luhlola kokubili inani lobuhle kanye nekhwalithi yokukhiqiza nge-taxonomy equkethe ubukhulu obuncane obungama-52.
Iphepha, elithunyelwe ku-arXiv ngo-Aug. 26, lethula i-VGA-BenchV2 njengesandiso sebhentshimakhi yangaphambili ebizwa nge-VGA-Bench. Ababhali bagcina izilinganiso ezimbili eziyinhloko—i-Aesthetic ne-Generation—futhi bazihlukanise zibe izingxenye ezincane ezingama-52. Umgomo oshiwo uwukuhlola ikhwalithi yokukhiqizwa kwevidiyo ngezinga elingcono kakhulu kunephuzu elilodwa lilonke, kuyilapho futhi kusekela ukuthuthukiswa kwakamuva kwamajeneretha ahlolwayo. Ngakho-ke incazelo yethula ibhentshimakhi njengokubili uhlelo lokuhlola oluningiliziwe kanye nesisekelo sokwenza kangcono kamuva.
Ibhentshimakhi isebenzisa ukwaziswa okuhlukahlukene okungu-1,016 namavidiyo angaphezu kuka-60,000 akhiqizwa amamodeli esizukulwane samavidiyo ayi-12 avamile. Ababhali bathi ibhentshimakhi enwetshiwe ingeza izichasiselo zabantu ezisezingeni lomsebenzi ezingama-36,000: eziyi-16,200 zekhwalithi yobuhle, eziyi-13,200 zokumaka kobuhle kanye nezingu-6,600 zekhwalithi yokukhiqiza. Bachaza lokhu njengokukhuphuka kuka-VGA-Bench izikhathi ezingu-13.46, izikhathi ezingu-11.15 kanye nezikhathi ezingu-1.55, ngokulandelana. Lezi zibalo zibikwa njengengxenye yencazelo yephepha yesilinganiso nokunwetshwa kwezichasiselo.
I-VGA-BenchV2 ihlanganisa izingxenye ezintathu zomhloli. I-VAQA-Net isetshenziselwa ukuthola amaphuzu obuhle obuqhubekayo, kuyilapho i-VTag-Net ne-VGQA-Net ichazwa njengabahloli bemodeli enkulu yombono-yolimi olusekelwe ku-Qwen ukuze kufakwe ukumaka kobuhle nokuhlola ikhwalithi yesizukulwane. Leli phepha liphinde lethule ipayipi lokuhlola ukuze kwenziwe ngcono lapho umhloli ofundile wobuhle esebenza njengemodeli yomvuzo yokulungisa kahle okusekelwe ekuqiniseni ukufunda kwamajeneretha wevidiyo. Ababhali babika ukuqondana okuqinile nokwahlulela kwabantu kuwo wonke amamodeli esizukulwane abawahlolile, kodwa umthombo awunikezi amaphuzu ayisisekelo, iziqhathaniso noma imininingwane yokuhlola ku-abstract. Ngamanye amazwi, umhloli akachazwa nje kuphela njengethuluzi lokulinganisa, kodwa futhi njengengxenye yokugeleza komsebenzi okuhlongozwayo.
Imininingwane yomthombo: arxiv.org ↗
Kungani kubalulekile
Umsebenzi uhlose ukwenza ukuhlolwa kwesizukulwane sevidiyo kubonise eduze kakhulu ukwahlulela kwabantu ngenkathi kuxhunywa ukuhlola ekuthuthukisweni kwemodeli. Uma ukuqondanisa kwayo okubikiwe kanye nemiphumela yokuthuthukisa ivamile, uhlaka lunganikeza onjiniyela indlela efanayo yokuqhathanisa amajeneretha nokuthuthukisa ikhwalithi yokubuka.
Ividiyo ekhiqizwe yi-AI yahlulelwa ngaphezu kokuthi ukulandelana kukhiqizwa ngobuchwepheshe. Izibukeli zingase futhi zikhathalele ukwakheka, ukukhanga okubonakalayo kanye nokuthi umnyakazo okhiqiziwe noma okuqukethwe kuyawanelisa yini ukwaziswa. Isakhiwo esishiwo se-VGA-BenchV2 sibalulekile ngoba sihlukanisa ukuhlolwa kobuhle kusukela kukhwalithi yesizukulwane futhi sihlukanisa kokubili kobukhulu obuncane, okungenzeka senze ubuthakathaka bubonakale kalula kunesilinganiso esisodwa esihlanganisiwe. Lokho kuhlukaniswa kuwukukhetha kwenhlangano okumaphakathi okuchazwe kubhentshimakhi.
Isikali sedathasethi ebikiwe singenza ibhentshimakhi ibe usizo njengengqalasizinda eyabiwe yocwaningo. Iqoqo elihlanganisa ukwaziswa okungu-1,016, amavidiyo akhiqiziwe angaphezu kuka-60,000 namamodeli angu-12 linikeza ababhali isisekelo sokuqhathanisa amasistimu kukho konke okokufaka okuhlukahlukene kunokuthembela kusethi encane yemiboniso. Izichasiselo zomuntu zibaluleke kakhulu enjongweni eshiwo yephepha: umhloli uqeqeshwe ngokumelene nezahlulelo ababhali abahlose ukubonisa okuthandwa abantu, kunokuncika kuphela ezilinganisweni ezizenzakalelayo. Iphepha lethula leli qoqo njengesisekelo sokuqeqeshwa kokuqhathanisa nokuhlaziya.
Isici esibaluleke kakhulu isixhumanisi phakathi kokulinganisa nokwenza kahle. Ngokusho kwephepha, isihloli sobuhle singasetshenziswa njengemodeli yomvuzo ngesikhathi sokuqinisa-ukufunda kahle, okuvumela ijeneretha ukuthi ilungiselele izimfanelo izilinganiso zebhentshimakhi. Lokho kungase kufinyeze indlela esuka ekukhombeni ubuthakathaka kuye ekuzameni ukuthuthukiswa okuhlosiwe. Nokho, umthombo usungula lokhu njengohlaka olubikiwe kanye nomphumela wokuhlola, hhayi njengobufakazi bokuthi le ndlela isebenza ngokuthembekile ekukhiqizweni noma ithuthukisa zonke izici zekhwalithi yevidiyo. Lokhu kuhlobana kwephepha phakathi kokuhlola nokuthuthukiswa kwemodeli.
Umsebenzi ungase futhi ube nomthelela ekutheni amasistimu okukhiqiza amavidiyo wesikhathi esizayo aqhathaniswa kanjani. Uhlaka lokuhlola oluvamile lungenza izimangalo ezimayelana nenqubekela phambili zichazeke kalula uma abacwaningi abahlukene basebenzisa ukwaziswa okufanayo, ubukhulu kanye nezinyathelo eziqondaniswe nomuntu. Lokho kubaluleka kuhlala kunemibandela: i-abstract ayisunguli ukuqinisekiswa okuzimele, ukutholwa amaqembu angaphandle, ukukhiqizwa kabusha kuwo wonke amasethi wedatha noma ukuthi incazelo yebhentshimakhi yenani lobuhle imele izethameli ezinokukhethwa okuhlukile. Le mikhawulo ichaza ukuthi yini engaphethwa emthonjeni njengoba kunikeziwe.
I-Interactive Mechanism: Indlela Esebenza Ngayo Ngempela
Hlola ubuchwepheshe obuyisisekelo ngemuva kwalokhu kuthuthukiswa ngokuhlanganyela.
crm_get_transaction(id='4092').Which component of an AI application is the machine-learning model itself?
Ongakubuka ngokulandelayo
Imibuzo ebalulekile evulekile eyokuthi ingabe abahloli bahlala bethembekile yini ngaphezu kwamamodeli ayi-12 kanye nesethi yokwaziswa esetshenziswe ephepheni, noma ngabe ukuthuthukiswa kokufunda okuqiniswayo kudluliselwe ekwazisweni namamodeli angabonakali, nokuthi ingabe ukulungiselelwa okwengeziwe kubhentshimark kudala izintandokazi ezibukwayo eziwumngcingo noma ezingafuneki.
Umbuzo wokuqala uwukuthi jikelele. Izivivinyo ezibikiwe zihlanganisa amamodeli esizukulwane samavidiyo angu-12 kanye nomklamo webhentshimakhi kanye nesichasiselo, kodwa umthombo awusho ukuthi abahloli basebenza kanjani kumamodeli akamuva, izizinda ezingabonwa, ukwaziswa okungajwayelekile noma amavidiyo anezici ezingekho eqoqweni. Ukuhlola okuzimele kungasiza ekunqumeni ukuthi ukuqondanisa komuntu okubikiwe kubanzi noma kuhlobene eduze nokwakhiwa kwebhentshimakhi. Okwamanje, ububanzi balokho kuqondaniswa okubikiwe buhlala buvulekile.
Umbuzo wesibili ukuthi ingabe ukulungiselelwa okuqondiswa kwebhentshimakhi kukhiqiza ukuthuthukiswa okuhlala isikhathi eside. Iphepha lithi umhloli wobuhle usetshenziswa njengemodeli yokuklomelisa yokuqinisa ukufunda kahle, kodwa umthombo awunikezi usayizi wezinzuzo, izindleko zokuqeqesha, ukuhlukaniswa kokuhlola, noma ubufakazi bokuthi ukuthuthukiswa kokudluliswa kwebhentshimakhi ezahlulelweni zabantu ngaphandle kwayo. Leyo mininingwane iyadingeka ukuze kuhlolwe ukubaluleka okungokoqobo. Ziphinde zishiye ukubaluleka okungokoqobo komphumela wokulungiselela kungaxazululiwe.
Ingcuphe ehlobene iwukuthuthukisa kulokho okungalinganiswa. Uma ijeneretha icushwe nomhloli ofundile, ingathuthukisa isikolo sayo kuyilapho ingahluka kancane, ithembeke kancane ekwazisweni noma ingakhangi kancane kubantu abancamelayo yabo imelelwe kancane ezichasiselweni. I-abstract ayikubiki ukwehluleka okunjalo, ngakho-ke kufanele kuthathwe njengombuzo wokuhlola ongaxazululiwe kunokutholwa kwephepha. Lokho kungenzeka yingakho ukuziphatha komhloli ngemuva kokwenza kahle kusasele kufanele kubhekwe.
Abafundi kufanele futhi babuke izinsiza zephepha nezinto zobuchwepheshe, irekhodi le-arXiv elithi zitholakala ngesigaba sensiza esixhunyiwe. Umthombo onikezwe lapha awukucacisi ngqo okuqukethwe kwalezo zinsiza, imigomo yelayisense noma isimo sokukhishwa. Ukukhiqizwa kabusha ngabacwaningi kusetshenziswa idatha, ukwaziswa, izichasiselo nezingxenye zomhloli kuzocacisa ukuthi ingakanani i-VGA-BenchV2 engasebenza njengezinga eliwusizo elibanzi kunokuba njengomphumela oboshelwe ocwaningweni olulodwa. Leyo mibuzo iphathelene nobubanzi obusebenzayo bokukhishwa, hhayi umphumela obikiwe ku-abstract.