12. Min-Max Game
● Generator trying to Fool the discriminator
● Discriminator somehow needs to identify fake from real very well
● Something similar to min-max in game theory
Since it’s Adversary
26. Main Problem - Discriminator Saturation
● Discriminator is too Good :(
● There won’t be any chance for the generator to learn something
Yunjey Choi
30. ● GAN's Task is to make the Generated Distribution(Pmodel) same as the Real data
Distribution(Preal)
31. ● There are ways to measure similarity of two distributions Eg:
○ KL divergence
○ Jensen–Shannon divergence
We can easily prove that Optimization of GAN’s loss function is similar to reducing Jensen–Shannon
divergence between the two distributions
32.
33. When we have an optimal discriminator
Optimization of the Loss = Minimizing the Jensen Shannon Divergence
35. GAN is not easy to Train !
● Non-convergence: the parameters oscillate, constantly destabilize and unlikely to arrive to
converge (Issues with Nash Equality).
● Mode collapse: generator collapses, leading to produce limited varieties of samples.
36.
37. Yes! there are more stable methods right now !
❖ Wasserstein GAN
WGAN vs GAN - Similar in terms of Formality & Functionality
Only thing change is the Loss Function !
38. Now Loss Function is more of a Critic !
❖ Previously the Discriminator and the Generator are working against each
other
❖ But now discriminator is is trying to give the generator an Idea of how different
it’s generated data is deviate from the actual data distribution.
❖ No Log probabilities - No Diminish Gradients
❖ Uses EM(Earth Mover's Distance) distance to model the loss function !
39. Wasserstein Distance or EM Distance
This is a measurement about how much work that generator has to do to match the
distribution of the real images
This is why we call it a Critic!
40. Reducing the distance between generated samples and real samples
Generator
distribution
Real
distribution
Critic
44. We need to clip the weights in the discriminator
● f has to be a 1-Lipschitz function.
● To enforce the constraint, WGAN applies a very simple clipping to restrict the
maximum weight value in f
● The weights of the discriminator must be within a certain range controlled by the
hyperparameters
After every update we need to clip the weights 0f the discriminator
49. Resources
GAN - https://arxiv.org/abs/1406.2661
WGAN - https://arxiv.org/abs/1701.07875
Improved WGAN - https://arxiv.org/abs/1704.00028
Principal Method Of Training GAN - https://openreview.net/pdf?id=Hk4_qw5xe
Amazing series of Article By Jonathan Hui
https://medium.com/@jonathan_hui/gan-whats-generative-adversarial-networks-and-its-application-f39ed278ef09
50. What we are into this ..
❖ GANhas an amazing ability to enrich Reinforcement Learning such as…
1. Planning
2. Inverse Reinforcement
51. Imitation Learning
● Learning From Expert’s Demonstrations
● Something in between Supervised Learning and Deep Reinforcement Learning
● There is a clear connection between GAN and Imitation Learning
53. What is the difference !
1. Instead of just Images we have expert’s trajectories which means states and
action pairs
2. Now the Generator is an AI Agent