Cibiyoyin Sadarwar Masu Adawa na Halitta (GANs), wanda Ian Goodfellow da sauransu suka gabatar a shekara ta 2014, suna wakiltar sauyi a cikin koyon da ba a kula da shi ba da kuma rabin kulawa. Wannan tsari yana sanya cibiyoyin sadarwar jijiyoyi guda biyu—Mai Halitta da Mai Rarrabewa—a kan juna a cikin wasan minimax. Babban manufar ita ce koyon samar da sabbin bayanai waɗanda ba za a iya bambanta su da ainihin bayanai ba. Wannan takarda tana ba da cikakken bincike kan tsarin GAN, ƙalubalen horar da su, hanyoyin kimantawa, da hangen nesa na gaba akan ci gabansu da aikace-aikacensu.
Ƙirar GAN ta asali ta kafa ƙa'idar horarwa ta adawa wacce ke goyan bayan duk bambance-bambancen da ke biyo baya.
Tsarin ya ƙunshi sassa biyu:
An tsara horarwa azaman wasan minimax na ƴan wasa biyu tare da aikin ƙima V(G, D):
$\min_G \max_D V(D, G) = \mathbb{E}_{x \sim p_{data}(x)}[\log D(x)] + \mathbb{E}_{z \sim p_z(z)}[\log (1 - D(G(z)))]$
A aikace, horarwa tana musanya tsakanin inganta D don haɓaka daidaiton rarrabawarsa da inganta G don rage $\log(1 - D(G(z)))$. Ƙalubalen gama gari sun haɗa da rugujewar yanayi, inda G ke samar da iyakantattun nau'ikan samfura, da rashin kwanciyar hankali na horarwa.
Don magance iyakokin tushe, an gabatar da tsare-tsare masu yawa na ci gaba.
cGANs, wanda Mirza da Osindero suka gabatar, sun faɗaɗa tsarin asali ta hanyar sanya sharadi akan duka mai halitta da mai rarrabewa akan ƙarin bayani y (misali, alamomin aji, bayanin rubutu). Wannan yana ba da damar sarrafa samar da takamaiman nau'ikan bayanai. Aikin manufa ya zama:
$\min_G \max_D V(D, G) = \mathbb{E}_{x \sim p_{data}(x)}[\log D(x|y)] + \mathbb{E}_{z \sim p_z(z)}[\log (1 - D(G(z|y)))]$
Cibiyoyin Sadarwar Masu Adawa Masu Daidaituwar Zagaye (CycleGAN), wanda Zhu da sauransu suka gabatar, suna ba da damar fassara hoto zuwa hoto ba tare da bayanan horarwa biyu ba. Yana amfani da nau'i-nau'i biyu na mai halitta-mai rarrabewa kuma yana gabatar da asarar daidaiton zagaye don tabbatar da cewa fassara hoto daga yanki A zuwa B sannan komawa A yana samar da ainihin hoton. Wannan ya kasance abin tarihi ga fassarar yanki mara biyu, kamar yadda aka yi cikakken bayani a cikin takardarsu ta asali.
Kimanta GANs da ƙididdiga ba abu ne mai sauƙi ba. Ma'auni na gama gari sun haɗa da:
Asarar adawa ita ce ginshiƙi. Mafi kyawun mai rarrabewa don ƙayyadaddun mai halitta ana bayar da shi ta:
$D^*(x) = \frac{p_{data}(x)}{p_{data}(x) + p_g(x)}$
Sauya wannan cikin aikin ƙima yana nuna cewa ana samun mafi ƙarancin ƙa'idar horarwa ta zahiri lokacin da $p_g = p_{data}$, kuma ƙimar ita ce $-\log 4$. Ana iya ganin tsarin horarwa azaman rage bambancin Jensen-Shannon (JS) tsakanin rarrabawar bayanai na ainihi da na halitta, ko da yake aikin daga baya ya gano iyakokin bambancin JS, wanda ya haifar da madadin kamar tazarar Wasserstein da aka yi amfani da ita a cikin WGANs.
GANs na zamani kamar StyleGAN2 da BigGAN suna nuna sakamako mai ban mamaki. A kan bayanan kamar FFHQ (Flickr-Faces-HQ) da ImageNet:
Bayanin Ginshiƙi: Zanen ginshiƙi na hasashe zai nuna ci gaban makin FID akan lokaci don samfura kamar DCGAN, WGAN-GP, StyleGAN, da StyleGAN2 akan bayanan CelebA, yana nuna bayyanannen yanayin raguwa (ingantawa) a cikin FID, yana nuna saurin ci gaba a cikin ingancin samarwa.
Tsarin don Kimanta Sabuwar Takardar GAN:
Nazarin Lamari: Nazarin GAN na Rubutu-zuwa-Hoto: Aiwatar da tsarin. Samfurin yana amfani da mai rikodin rubutu na tushen transformer da mai halitta na StyleGAN2. Ƙirar ta ta'allaka ne akan kulawar tsaka-tsaki. Yana yiwuwa yana amfani da asarar kwatance tare da asarar adawa. Duba FID akan bayanan COCO ko CUB akan ma'auni kamar AttnGAN ko DM-GAN. Kimanta ko takardar ta haɗa da nazarin cirewa wanda ke tabbatar da gudunmawar kowane sabon ɓangare.
Hanyar ci gaban GAN tana nuni zuwa ga yankuna masu mahimmanci da yawa:
Hankali na Asali: Juyin juya halin GAN ba game da "aikace-aikace mai kisa" guda ɗaya ba ne, amma game da kafa koyon adawa a matsayin mahimmin, sassauƙan fifiko don kimanta yawa da haɗakar bayanai. Ƙimar sa ta gaskiya tana cikin samar da tsari inda "mai rarrabewa" zai iya zama kowane ma'auni na gaskiya wanda za a iya bambanta shi, yana buɗe kofofi nesa da samar da hoto—daga ƙirar kwayoyin halitta zuwa simintin kimiyyar lissafi, kamar yadda ake gani a ayyuka a DeepMind da kamfanonin AI daban-daban na biotech.
Kwararar Hankali & Juyin Halitta: Labarin yana bayyananne: daga wasan minimax na tushe (Goodfellow da sauransu), fagen ya yi sauri ya rabu don magance kurakurai nan take. cGANs sun ƙara sarrafawa. WGANs sun kai hari rashin kwanciyar hankali ta hanyar kafa asarar a cikin tazarar Wasserstein bisa ka'ida. StyleGANs sun raba wuraren ɓoye don sarrafawa da ba a taɓa gani ba. CycleGAN ya magance matsalar bayanan biyu. Kowane mataki ba kawai ingantacciyar ci gaba ba ne; juyawa ce ta dabarun da ke magance raunin gindi, yana nuna fagen da ke maimaitawa cikin sauri mai karye.
Ƙarfi & Kurakurai: Ƙarfin ba shakku ne: ingancin fitarwa mara misaltuwa a cikin yankuna kamar hoto da sauti. Mai sukar adawa aikin asara ne mai ƙarfi, wanda aka koya. Duk da haka, kurakurai na tsarin ne. Horarwa ya kasance sanannen rashin kwanciyar hankali kuma yana da hankali ga hyperparameters—"baƙar fata." Rugujewar yanayi fatalwa ce mai dagewa. Kimantawa har yanzu batu ne mai kaifi; ma'auni kamar FID wakilai ne, ba cikakkun ma'auni na amfani ba. Bugu da ƙari, kuɗin lissafi don samfuran SOTA yana ban mamaki, yana haifar da shinge ga shiga kuma yana ɗaga damuwar muhalli.
Hankali Mai Aiki: Ga masu aiki: Kada ku fara daga GANs na vanilla. Gina akan tsare-tsare masu kwanciyar hankali kamar StyleGAN2/3 ko amfani da bambancin asarar Wasserstein daga ranar farko. Ba da fifiko ga ingantaccen kimantawa ta amfani da ma'auni da yawa (FID, Daidaici/Tunawa). Ga masu bincike: 'Ya'yan itace masu ƙanƙanta sun tafi. Gaba gaba ba kawai mafi kyawun hotuna ba ne, amma inganta inganci, sarrafawa, da aikace-aikace ga bayanan da ba na gani ba. Bincika samfuran gauraye; hawan Samfuran Watsawa yana nuna cewa horarwa ta adawa ba ita kaɗai ce hanyar inganci ba. Gaba ba na GANs kaɗai ba ne, amma ga tsare-tsare masu ka'ida waɗanda za su iya amfani da horarwa mai kwanciyar hankali, ɓoyayyun bayanai masu fassara, da samfurin inganci—GANs na iya zama muhimmin sashi, amma mai yiwuwa ba tsarin gini kaɗai ba.