The document discusses Generative Adversarial Networks (GANs), a type of generative model proposed by Ian Goodfellow in 2014. GANs use two neural networks, a generator and discriminator, that compete against each other. The generator produces synthetic data to fool the discriminator, while the discriminator learns to distinguish real from synthetic data. GANs have been used successfully to generate realistic images when trained on large datasets. Examples mentioned include Pix2Pix for image-to-image translation and STACKGAN for text-to-image generation.
3. Generative Model 1D example
[0, 1]
http://m.blog.naver.com/atelierjpro/22098475
8512
4. Generative Adversarial Nets
GAN started by Ian Goodfellow
[https://arxiv.org/abs/1406.2661]
GAN used for generating realistic data
( usually, for images )
7. GAN : How does it works?
Random NoiseInput : Random Noise
Output: Realistic Image
Discriminator
Network
8. GAN : How does it works?
< Traditional Training Model >
9. GAN : How does it works?
< GAN Training Model > MAIN IDEA : Just showing lot’s of images
Initial
Trained
10. GAN : How does it works?
< GAN Training Model > MAIN IDEA : Just showing lot’s of images
11. Generative Model 1D example
http://m.blog.naver.com/atelierjpro/22098475
8512
12. Generative Model 1D example
http://m.blog.naver.com/atelierjpro/22098475
8512
g = G(z) # z is uniform distribution
Generative Function
Input : Uniform distribution
Output : Real Data
13. Generative Model 1D example
http://m.blog.naver.com/atelierjpro/22098475
8512
g = G(z) # z is uniform distribution
Generative Function : G is neural network
Input : Uniform distribution
Output : Real Data
Not Trained Model
(yellow)
14. Generative Model 1D example
http://m.blog.naver.com/atelierjpro/22098475
8512
(0 || 1) = D(d) # d is data ( from gen func or real data )
Discriminator : D is neural network
Input : Data
Output : real or fake
Black is D ( input from not trained G )
Yellow is G ( not trained ) - 1
Red is input of G
Blue is Real Data
Discriminator가 잘 구분함
( G와 Real이 완전이 분리)
15. Generative Model 1D example
http://m.blog.naver.com/atelierjpro/22098475
8512
(0 || 1) = D(d) # d is data ( from gen func or real data )
Discriminator : D is neural network
Input : Data
Output : real or fake
Black is D ( input from not trained G )
Yellow is G ( not trained ) - 2
Red is input of G
Blue is Real Data
Discriminator가 잘 구분 못함
( G와 Real이 섞여있음)
21. Apple’s First Paper
Learning from Simulated and Unsupervised Images through Adversarial Training
https://arxiv.org/abs/1612.07828
Image to Image
Fake to REAL