Discriminator — Generative Adversarial Network’da input sample real training data’danmi yoki generator yaratgan synthetic distributiondanmi, shuni ajratishga o‘rgatiladigan neural network. U generator uchun o‘rganiluvchi sifat signali yaratadi. Ayrim GAN variantlarida output probability emas, cheklanmagan scalar score bo‘ladi va model critic deb ataladi.
Classification vazifasi
Oddiy discriminator real sample’ga 1, fake sample’ga 0 target bilan binary classification loss oladi. Generator sample’i discriminator orqali o‘tganda gradient generatorgacha qaytadi. Shu sabab discriminator faqat yakuniy hakam emas; uning learned feature va decision boundarysi generator qaysi xatoni tuzatishini belgilaydi.
Training batch real va fake sample’lardan tuziladi. Fake sample generatorning joriy versiyasidan olinadi va discriminator yangilanishida ko‘pincha generator graphidan detach qilinadi. Keyingi generator qadamida yangi fake sample yaratiladi, discriminator parameterlari yangilanmaydi, ammo input bo‘yicha gradient o‘tadi.
Arxitektura tanlovi
Image discriminator convolution bilan local texture va global structure’ni tekshiradi. PatchGAN butun rasmga bitta score berish o‘rniga patch grid chiqaradi; bu texture aniqligiga foydali, lekin global geometry’ni alohida constraint talab qilishi mumkin. Multi-scale discriminator turli resolutionda ishlaydi.
Audio discriminator waveformni bir necha sample rate va period bo‘yicha ko‘radi. Period discriminator voiced signal harmonikasini, scale discriminator transient va spectral envelope’ni ushlaydi. Conditional discriminator sample bilan birga class label, text embedding yoki source image’ni qabul qilib, faqat realness emas, condition mosligini ham tekshiradi.
Muvozanat va regularization
Discriminator training data’ni yodlab olsa real va fake’ni juda oson ajratadi, generatorga gradient sifati pasayadi. Data augmentation, dropout, weight decay va limited capacity overfittingni kamaytiradi. Spectral normalization layer weightining eng katta singular qiymatini nazorat qilib Lipschitz behaviorni yaxshilaydi.
Gradient penalty real va fake orasidagi interpolated sample’da gradient normini targetga yaqinlashtiradi. R1 regularization real data bo‘yicha input gradientini jazolaydi. Update ratio discriminator va generator tezligini muvozanatlashtiradi. Loss curve’dan tashqari real/fake accuracy va gradient norm kuzatiladi, ammo ideal accuracy uchun universal qiymat yo‘q.
Critic tushunchasi
Wasserstein GAN’dagi critic probability bermaydi va sigmoid bilan cheklanmaydi. U real sample’ga yuqoriroq, fake sample’ga pastroq score berib Wasserstein distance taxminini yaratadi. Uni oddiy classifier probabilitysi sifatida talqin qilish xato. Lipschitz constraint weight clipping, gradient penalty yoki spectral normalization bilan ta’minlanadi.
Energy-based yoki hinge-loss GANlarda ham outputning ma’nosi objectivega bog‘liq. Shuning uchun implementationda activation va loss juftligi mos bo‘lishi kerak. Binary cross-entropy with logits ishlatilsa model oxirida alohida sigmoid kerak emas; ikki marta sigmoid gradientni buzishi mumkin.
Baholashdagi o‘rni
Training discriminatorini keyinchalik universal deepfake detector sifatida ishlatib bo‘lmaydi. U aynan joriy generator va dataset artefactlariga moslashgan. Yangi generator, compression yoki domain shift’da accuracy keskin tushishi mumkin. Detection uchun mustaqil data, calibration va adversarial robustness bahosi kerak.
Discriminator feature’lari perceptual loss yoki representation sifatida ishlatilishi mumkin, lekin training objective bias’ini olib yuradi. Generator sifatini FID, diversity va human evaluation bilan alohida o‘lchash zarur. Identity yoki maxfiy data bilan o‘qitilganda real sample access’i va retentioni ham himoyalanadi.
Bog‘liq tushunchalar
Generative Adversarial Network, Generator, Critic, Binary classification, Gradient penalty, Spectral normalization, PatchGAN