GAN

GAN — ikkita neural networkning raqobatli trainingi orqali yangi data namunalarini yaratishga mo‘ljallangan generativ model architecture’si. To‘liq nomi Generative Adversarial Network. GAN odatda generator va discriminator deb ataladigan ikki modeldan tashkil topadi.

Generator haqiqiy dataga o‘xshash namuna yaratadi, discriminator esa namuna real datasetdanmi yoki generator tomonidan yaratilganmi, shuni ajratishga harakat qiladi.

Generator

Generator random latent vectorni input sifatida oladi:

z → generator → synthetic sample

Output:

• rasm;
• audio;
• signal;
• boshqa data

bo‘lishi mumkin.

Training davomida generator discriminatorni aldashga yaxshiroq o‘rganadi.

Discriminator

Discriminator real va generated namunani qabul qilib, ularning manbasini klassifikatsiya qiladi.

Uning vazifasi:

real → yuqori real score
fake → past real score

berish.

Discriminator generatorga gradient signal beradi.

Adversarial training

Ikki model qarama-qarshi objective bilan o‘qitiladi:

• discriminator real va fakeni ajratishni yaxshilaydi;
• generator discriminatorni aldashni yaxshilaydi.

Bu minimax game sifatida ifodalanadi.

Training muvozanati buzilsa bir model ikkinchisidan ancha kuchli bo‘lib qolishi mumkin.

Latent space

Generator inputidagi random vector latent space nuqtasi hisoblanadi.

Latent dimension outputning yashirin factorlarini ifodalashi mumkin.

Yaqin latent nuqtalar o‘xshash namuna yaratishi ehtimoli bor.

Latent interpolation ikki generated sample orasidagi silliq o‘tishni ko‘rsatadi.

Training data

GAN real distributionni training datasetdan o‘rganadi.

Dataset:

• yetarli katta;
• sifatli;
• representativ;
• bir xil preprocessing;
• huquqiy jihatdan foydalanishga mos

bo‘lishi kerak.

Dataset bias generated outputda ham ko‘rinadi.

Loss

Original GAN binary classificationga yaqin loss ishlatadi.

Generator va discriminator objective’lari turli formulada yozilishi mumkin.

Amaliy variantlar:

• non-saturating loss;
• Wasserstein loss;
• hinge loss;
• least-squares loss.

Loss qiymati visual sifat bilan har doim to‘g‘ridan-to‘g‘ri bog‘liq emas.

Mode collapse

Generator turli inputlarga juda o‘xshash yoki cheklangan turdagi output yaratib qolishi mumkin.

Bu mode collapse.

Masalan, datasetda ko‘p yuz turi bo‘lsa ham generator faqat bir necha ko‘rinishni takrorlaydi.

Diversity metric va sample inspection muhim.

Training instability

GAN training ko‘pincha nozik:

• learning rate;
• optimizer;
• normalization;
• architecture;
• update nisbati;
• batch size;
• discriminator strength

natijaga ta’sir qiladi.

Loss oscillation va gradient muammolari yuz berishi mumkin.

DCGAN

Deep Convolutional GAN rasm generation uchun convolutional networklardan foydalanadi.

Generator upsampling yoki transposed convolution bilan rasm o‘lchamini oshiradi.

Discriminator convolution orqali rasm feature’larini ajratadi.

DCGAN GAN architecture’sining muhim klassik variantlaridan biri.

Conditional GAN

Generator va discriminatorga qo‘shimcha label yoki condition beriladi.

Masalan:

class = "mushuk"

Shunda model aynan shu classga mos rasm yaratadi.

Condition text, category, segmentation map yoki boshqa signal bo‘lishi mumkin.

Image-to-image

GAN bir rasm domainidan boshqasiga transform qilish uchun ishlatilishi mumkin.

Misollar:

• sketch → photo;
• day → night;
• mask → image;
• low resolution → high resolution;
• old photo → restored photo.

Paired yoki unpaired dataset yondashuvga bog‘liq.

Super-resolution

Past resolution rasmni yuqori resolutionga aylantirishda GAN perceptual detail yaratishi mumkin.

Generated mayda detail real source’da mavjud bo‘lmasligi ehtimoli bor.

Tibbiy yoki forensic tasvirda bunday “ishonarli, ammo uydirma” detail xavfli.

Style transfer

GAN ma’lum visual domain yoki uslubni o‘rganib, contentni shu style’ga aylantirishi mumkin.

Identity, geometry va rangni qanchalik saqlash objective hamda datasetga bog‘liq.

Artistic outputda subjective evaluation muhim.

Evaluation

Generative sifatni baholash usullari:

• FID;
• Inception Score;
• precision va recallga oid generative metric;
• diversity;
• human evaluation;
• task-specific score.

Metric modelning barcha sifatini to‘liq ko‘rsatmaydi.

GAN va VAE

VAE probabilistik latent model va reconstruction objective’ga tayanadi.

GAN adversarial discriminator orqali keskin va realistik output yaratishga intiladi.

VAE latent space va likelihoodga yaqin modelda barqarorroq, GAN esa visual sharpnessda kuchli bo‘lishi mumkin.

GAN va Diffusion

Diffusion model shovqindan bosqichma-bosqich sample yaratadi.

U training stability va diversity bo‘yicha ko‘p vazifalarda GANdan ustunlashgan.

GAN inference’da bir forward pass bilan tez output yaratishi mumkin.

Architecture tanlovi latency va sifatga bog‘liq.

Deepfake

GAN real inson yuzi yoki ovoziga o‘xshash synthetic media yaratishda ishlatilgan.

Bu:

• kino;
• dubbing;
• kreativ tool

uchun foydali bo‘lishi mumkin.

Shuningdek impersonation, misinformation va fraud xavfini yaratadi.

Consent va provenance zarur.

Data augmentation

GAN kam uchraydigan class uchun synthetic sample yaratishda ishlatilishi mumkin. Generated data real datasetni to‘liq almashtirmaydi. Model trainingda artifactni o‘rganib, testda yomon natija berishi mumkin. Synthetic va real sample nisbati, diversity va label aniqligi tekshiriladi.

Memorization

Generator training rasmlarini aynan yoki juda yaqin ko‘rinishda qayta yaratishi ehtimoli bor. Nearest-neighbor tahlili va privacy testlari memorizationni baholaydi. Sensitive dataset uchun membership inference va model extraction xavfi ham ko‘riladi.

Checkpoint tanlash

GAN loss qiymati sample sifatini oddiy ko‘rsatmaydi. Training davomida checkpointlar FID, diversity va human inspection orqali solishtiriladi. Discriminator juda kuchli yoki zaif bo‘lgan nuqtalar alohida kuzatiladi.

Bog‘liq tushunchalar

Generative Adversarial Network, Generator, Discriminator, Adversarial training, Latent space, Mode collapse, DCGAN, Conditional GAN, Deepfake, Diffusion model