Taƙaitaccen Bayani

Cibiyoyin sadarwa masu hamayya (GANs) sun zama ingantacciyar hanyar da ba ta da kulawa don ƙirƙirar tsarin ƙididdiga na bayanan duniya na gaske, kamar hotuna na halitta. An horar da waɗannan hanyoyin sadarwa don taswirar shigarwar bazuwar a cikin sararin su zuwa sabbin samfura masu wakiltar bayanan da aka koya. Duk da haka, tsarin sararin ɓoyayye yana da wuya a fahimta saboda girman girmansa da rashin layi na janareta, yana iyakance amfanin samfuran.

Fahimtar sararin ɓoyayye yana buƙatar hanyar gano lambobin shigarwa don hotunan duniya na ainihi (juyawa), da kuma hanyar gano kwatance tare da sanannun sauye-sauyen hoto (fassara). Anan, muna amfani da tsarin jumulla don magance duka batutuwa guda biyu lokaci guda. Mun ƙirƙiri hanyar da ba ta da alaƙa da tsarin gine-ginen don lissafin ma'aunin Riemann na samfurin hoton da GANs suka ƙirƙira. Rarrabuwar halayen ma'auni yana ware gatari waɗanda ke lissafin matakan daban-daban na bambancin hoto.

Bincike na kwarewa na GANs da yawa da aka riga aka horar ya nuna cewa bambancin hoto a kusa da kowane matsayi yana mai da hankali ne kaɗan a kan manyan gatari kaɗan (sararin yana da girman anisotropic sosai) kuma kwatance waɗanda ke haifar da wannan babban bambanci suna kama a wurare daban-daban a cikin sarari (sarari yana da kama). Mun nuna cewa yawancin manyan ƙwayoyin eigenvectors sun dace da fassarorin da za a iya fassara a cikin sararin hoto, tare da wani ɓangare mai mahimmanci na eigenspace wanda yake dace da ƙananan sauye-sauye waɗanda za a iya matsawa su.

Wannan fahimtar jumullar ta haɗa mahimman sakamakon baya masu alaƙa da fassarar GAN. Mun nuna cewa amfani da wannan ma'auni yana ba da damar ingantaccen inganci a cikin sararin ɓoyayye (misali juyawar GAN) da kuma sauƙaƙe gano gatari masu fassara ba tare da kulawa ba. Sakamakonmu ya nuna cewa ayyana jumullar samfurin hoton GAN na iya zama tsari gabaɗaya don fahimtar GANs.

Gabatarwa

Samfuran ƙirƙira masu zurfi, musamman Cibiyoyin Sadarwa Masu Hamayya (GANs), sun kawo juyin juya hali a fannin koyo mara kulawa ta hanyar ba da damar samar da hotuna masu kama da na gaske da kuma bambancin su. Duk da babban nasarar da suka samu wajen samar da samfuran hoto na gaske, tsarin asali na sararinsu na ɓoyayye ya kasance ba a fahimta sosai ba. Girman girma, yanayin rashin layi na waɗannan wurare yana gabatar da manyan kalubale don fassara da aikace-aikace masu amfani.

Wannan takarda ta gabatar da hangen nesa na jumulla don nazari da fahimtar sararin ɓoyayye na GANs. Ta hanyar ɗaukar janareta a matsayin taswira mai santsi daga sararin ɓoyayye zuwa sararin hoto, zamu iya amfani da kayan aikin jumullar Riemann don siffanta tsarin samfurin hoton da aka samu. Wannan hanya tana ba da tsari guda ɗaya don magance manyan kalubale guda biyu a cikin binciken GAN: juyawar sararin ɓoyayye (nemo lambobin hotuna na gaske) da fassara (gano mahimman kwatance a cikin sararin ɓoyayye).

Ayyukanmu ya nuna cewa ma'aunin Riemann na samfurin GAN yana bayyana mahimman kaddarorin game da jumullarsa, gami da anisotropy da kama, waɗanda ke da tasiri kai tsaye ga duka fahimtar ka'ida da aikace-aikace masu amfani na samfuran ƙirƙira.

Bayanin Baya

Cibiyoyin sadarwa masu hamayya suna koyon alamu waɗanda ke siffanta rikitattun bayanai kuma daga baya suna samar da sabbin samfura masu wakiltar wannan saiti. A cikin 'yan shekarun nan, an sami babban nasara wajen horar da GANs don samar da hotuna masu ƙima da ƙima. GANs da aka horar da kyau suna nuna sassauƙan sauye-sauye tsakanin fitowar hoto yayin shiga tsakani a cikin sararin shigarwar su, wanda ke sa su zama masu amfani a aikace-aikace kamar gyaran hoto mai girma (canza halayen fuskoki), abu