3. • Discriminative vs. Generative
Generative Model
CNN classifier GAN
“cat”
an inverse function?
4. • Discriminative vs. Generative
Generative Model
https://duphan.wordpress.com/tag/generative-model/
data distribution
p(x,z) or p(x|z)
Difficult, but important
5. • Data distribution
Generative Model
A slide from “Crash Course on Machine Learning Part II”
주요 특징만 필요
세밀한 특성, 분포를
완전히 알아야함
6. • Applications?
• Structured prediction (e.g., output text)
• Much more robust prediction
• Anomaly detection
• …
Generative Model
10. • Variational Auto-Encoder
• Data의 generation은 잘 됨
• 그러나 blurry face
Why GANs are different?
“Variational Autoencoder and
Extensions”
by Aaron CouCourville
Sample
interpolation
11. • Why are generated samples blurry?
• Regress to the mean => blurry images
Why GANs are different?
Ledig, C. et al.
Photo-Realistic Single Image Super-Resolution Using a
Generative Adversarial Network.
MSE or Euclidean distance
12. Why GANs are different?
from “Tutorial on Theory and Application of Generative Adversarial Networks” (CVPR17)
13. • Difficult to hand-craft a good perceptual loss
function
• => 그럴듯한지 여부를 판단하는 neural network가
있다면?
Generative Adversarial Network
17. Training GANs
• Discriminator training
Deep Learning for Computer Vision: Generative models and adversarial training (UPC2016)
18. Training GANs
• Generator training
Deep Learning for Computer Vision: Generative models and adversarial training (UPC2016)
19. • Adversarial loss (vanilla GAN)
Training GANs
D Real sample은 D(x) -> 1 Generated sample은 D(x) -> 0
G Generated sample도 D(x) -> 1
20. • Iterate these two steps until convergence
• Eventually (we hope) that the generator gets so good that it
is impossible for the discriminator to tell the difference
between real and generated images. Discriminator
accuracy = 0.5
Training GANs
Discriminator
Data
Model
Distribution
Random guess
22. • In practice?
• Training stability
(D vs. G balance, …)
• Mode collapse
Training GANs
Generative Models II (CIFAR-CRM DLSS 2017) by Aaron Courville
An image from “Generative adversarial networks” by Namju Kim
24. • Realistic image!
GAN Results & Applications
Berthelot, D., Schumm, T., & Metz, L. (2017). BEGAN: Boundary
Equilibrium Generative Adversarial Networks.
realistic & diverse samples
Nguyen, A. et al. (2016). Plug & Play
Generative Networks: Conditional Iterative
Generation of Images in Latent Space
25. • Representation interpolation
GAN Results & Applications
Berthelot, D., Schumm, T., & Metz, L. (2017). BEGAN: Boundary
Equilibrium Generative Adversarial Networks.
Memorizing X
26. • Super-resolution
GAN Results & Applications
Ledig, C. et al. (2016) Photo-Realistic Single Image
Super-Resolution Using a Generative Adversarial
Network.
27. GAN Results & Applications
Low-res image
Generated
high-res image
Generated
or not?
28. • Image-to-image translation
GAN Results & Applications
Isola, P., Zhu, J.-Y., Zhou, T., & Efros, A. A. (2016). Image-to-
Image Translation with Conditional Adversarial Networks.
34. Literature Review (1) De-noising
Low-dose Routine-doseGAN-denoised
“Training with an adversarial network allows the generator to better learn the
noise distribution in routine-dose CT and produce more realistic images for
more accurate coronary calcium quantification.”
MSE-denoised
35. • CT => PET image translation
Literature Review (2) - Image-to-image translation
45. • Medical image generation from scratch?
• Structure + Dynamics + Variations + …
How to Tackle
46. • Medical image generation from scratch?
• Structure + Dynamics + Variations + …
• Might be too difficult
How to Tackle
(GAN은 global structure / counting 같은 것에서 특히 약함)
47. • (First step) image generation w/ structural hints
• Structure + Dynamics + Variations + …
• Image-to-image translation과 유사
How to Tackle
48. • Gaze estimation task
• Model-based synthetic data => GAN => Realistic
data
How to Tackle
50. How to Tackle
Self-regularization (minimizing feature dist.)
Local adversarial loss (=PatchGAN)
History of refined images
History 안쓰면 artifacts 발생
51. • Possible hints?
• Model-based synthetic data
• Normal image + conditioning
• Normal image + synthetic or expert-guided label
How to Tackle
Synthetic or guided label
Generator
52. First Trial
Generator
Discriminator
Real normal
Real patient
Image + label
(for controllability)
Generated patient
Our target
- synthetic or supervised label
- Matching image-segmentation pair?
- Realistic image?
54. • Typical GAN issues (training, …)
• Sample quality? (feat. M.D. researchers)
• How to evaluate generated samples
• Practical effectiveness?
• Unsupervised (or semi-supervised)
segmentation/classification in medical domain
Have far to go