Ababhali:
(1) Robert Honig, ETH Zurich ([email protected]);
(2) Javier Rando, ETH Zurich ([email protected]);
(3) uNicholas Carlini, i-Google DeepMind;
(4) UFlorian Tramer, ETH Zurich ([email protected]).
Isishwankathelo kunye 1. Intshayelelo
6.1 Iziphumo eziphambili: Lonke uKhuseleko lujikeleziswa ngokulula
Ingxoxo kunye neMpembelelo eBanzi, Imibulelo, kunye neeReferensi
A. Imizekelo yoBugcisa obuneenkcukacha
B. Robust Mimicry Izizukulwana
D. Umahluko ngeGlaze Finetuning
G. Iindlela zeSimbo sokuLingisa
H. IziKhuseli zeSimbo sokuLingisa esele zikhona
I. Iindlela zokulinganisa ezomeleleyo
Amagcisa axhalaba ngakumbi malunga nokuqhubela phambili kwimodeli yokuvelisa imifanekiso enokuthi iphindaphinde ngokusondeleyo iindlela zabo zobugcisa ezizodwa. Ukuphendula, izixhobo ezininzi zokukhusela ngokuchasene nesimbo sokulinganisa ziye zaphuhliswa ezibandakanya ukuphazamisana okuncinci kwimisebenzi yobugcisa epapashwe kwi-intanethi. Kulo msebenzi, sivavanya ukusebenza kokhuseleko oludumileyo-kunye nezigidi zokukhutshelwa-kwaye sibonisa ukuba zibonelela ngokhuseleko olungeyonyani. Sifumanisa ukuba umgudu ophantsi kunye neendlela "ezingekho kwi-shelf", ezifana nokunyuswa komfanekiso, zanele ukudala iindlela zokulinganisa ezinamandla ezithobisa kakhulu ukhuseleko olukhoyo. Ngophononongo lwabasebenzisi, sibonisa ukuba lonke ukhuselo olukhoyo lunokugqithwa ngokulula, lushiya amagcisa esesichengeni sokulinganisa isimbo. Siyalumkisa ukuba izixhobo ezisekelwe kwiingxaki ezichaseneyo azikwazi ukukhusela ngokuthembekileyo amagcisa ekusebenziseni kakubi i-AI yokuvelisa, kwaye sikhuthaza ukuphuhliswa kwezinye izisombululo zokukhusela.
Ukulinganisa isitayile sisicelo esidumileyo semifuziselo yesicatshulwa ukuya kumfanekiso. Kunikwe imifanekiso embalwa evela kumzobi, imodeli inokulungiswa ukuze kuveliswe imifanekiso emitsha ngolo hlobo (umzekelo, isiphekepheke kwisimbo sikaVan Gogh). Kodwa ukulinganisa isitayile kunokubangela ingozi enkulu ukuba kusetyenziswe kakubi. Ngokukodwa, amagcisa amaninzi anamhlanje anexhala lokuba abanye ngoku banokuvelisa imifanekiso ekopa isitayile sabo sobugcisa esahlukileyo, kwaye banokuba abathengi (Heikkila¨, 2022). Njengempendulo, ukhuseleko oluninzi luye lwaphuhliswa ukukhusela amagcisa kwisitayela sokulingisa (Shan et al., 2023a; Van Le et al., 2023; Liang et al., 2023). Olu khuselo longeza ukuphazamiseka okuchaseneyo kwimifanekiso epapashwa ngamagcisa kwi-intanethi, ukuze inyanzelise inkqubo yokulungiswa. Olu khuselo lufumene ingqwalasela ebalulekileyo kumajelo eendaba-kunye neempawu kwiNew York Times (Hill, 2023), CNN (Thorbecke, 2023) kunye neScientific American (Leffer, 2023)-kwaye zikhutshelwe ngamaxesha e-1M (Shan et al. , 2023a).
Nangona kunjalo, akukacaci ukuba ezi zixhobo ziwakhusela kangakanani amagcisa ngokuchasene nesimbo sokulingisa, ngakumbi ukuba umntu uzama ukubaphepha (Radiya-Dixit et al., 2021). Kulo msebenzi, sibonisa ukuba izixhobo zokukhusela isitayela se-state-of-the-art-Glaze (Shan et al., 2023a), Mist (Liang et al., 2023) kunye ne-Anti-DreamBooth (Van Le et al., 2023) -ayisebenzi xa ujongene neendlela ezilula zokulinganisa ezomeleleyo. Iindlela zokulinganisa ezinamandla esiziqwalaselayo zivela kwizicwangciso eziphantsi-ezifana nokusebenzisa iskripthi se-finetuning eyahlukileyo, okanye ukongeza ingxolo ye-Gaussian kwimifanekiso ngaphambi koqeqesho-kwizicwangciso zamanyathelo amaninzi adibanisa izixhobo ze-shelf. Siqinisekisa iziphumo zethu ngophando lomsebenzisi, olubonisa ukuba iindlela zokulinganisa ezinamandla zinokuvelisa iziphumo ezingabonakaliyo kumgangatho kwezo zifunyenwe kwimizobo engakhuselekanga (jonga uMzobo 1 womzekelo obonisa umzekeliso).
Sibonisa ukuba izixhobo ezikhoyo zokukhusela zibonelela ngombono wobuxoki wokhuseleko. Iindlela zethu zokulinganisa ezinamandla azifuni ukuphuhliswa kwezixhobo ezintsha okanye iindlela zokulungisa kakuhle, kodwa ngononophelo kuphela
ukudibanisa iindlela ezisemgangathweni zokusetyenzwa kwemifanekiso ezazisele zikho ngexesha ezi zixhobo zokhuseleko zaziqala ukwaziswa!. Ke ngoko, sikholelwa ukuba nabenzi bomgunyathi abanezakhono eziphantsi bebenokuziphepha ngokulula ezi zixhobo oko zasekwayo.
Nangona sivavanya izixhobo ezithile zokhuseleko ezikhoyo namhlanje, imida yokhuselo lokulinganisa isitayile. Amagcisa akwimeko engalunganga kuba kufuneka enze okuthile kuqala (oko kukuthi, xa umntu ekhuphele ubugcisa obukhuselweyo, ukhuselo alusenakuphinda lutshintshwe). Ukuze zisebenze, izixhobo ezikhuselayo zijongene nomsebenzi ocelomngeni wokudala izinto eziphazamisayo ezidlulisela kuyo nayiphi na indlela yokulungelelanisa, kwanabo bakhethwe ngokufanelekileyo kwixesha elizayo. Isigqibo esifanayo senziwe nguRadiya-Dixit et al. (Radiya-Dixit et al., 2021), owathi ukuxhatshazwa okuchaseneyo akukwazi ukukhusela abasebenzisi kwiinkqubo zokuqaphela ubuso. Ke ngoko siyalumkisa ukuba ubugcisa bokufunda ngomatshini ochaseneyo abuyi kukwazi ukukhusela amagcisa ngokuthembekileyo kukulinganisa isitayile sokuvelisa, kwaye sikhuthaze uphuhliso lwezinye iindlela zokukhusela amagcisa.
Sichaze iziphumo zethu kwizixhobo zokhuseleko ezichaphazelekayo phambi kokuba zipapashwe, ukuze zikwazi ukumisela eyona ndlela ilungileyo yokwenza kubasebenzisi abakhoyo.
Eli phepha li