gan-190318135433 (1).pptx
- 2. 2 • GANs were introduced in a paper by Ian Goodfellow and other researchers at the University of Montreal, including Yoshua Bengio, in 2014.
- 3. 3 –Yann LeCun “The most interesting idea in the last 10 years in Machine Learning”
- 4. 4 • So what are GANs? • What makes them so “interesting” ?
- 5. 5 • GAN - Generative Adversarial Network • GANs belong to the set of generative models. • It means that they are able to produce / generate new content.
- 6. Examples of results obtained with GANs Rightmost column contains true data that are nearest from the direct neighbouring generated samples 6
- 7. Generative model N = nxn grayscale image 7 Unrolled N-dim vector Problem is generating an N-dim vector that represents a “dog”
- 8. 8 • It is equivalent to generating a random variable with respect to the “dog probability distribution” • The “dog probability distribution” over the N dimensional vector space is a very complex one • We don’t know how to explicitly express this distribution • We don’t know how to directly generate complex random variables
- 9. 9 • It has been shown that Complex random variables can be built as a function of some simpler random variables C = F(S)
- 10. 10 • Uniform random variables can be generated easily • We need to express our N dim random variable as the result of a very complex function applied to a simple N dim random variable Complex r.v = Function(Simple r.v’s) • Solution - Use Neural network modelling
- 11. z ~ Pprior(z) 11 Generator function, G(z;θg) x ~ Pg(x;θg) Unrolled vector Generative Model
- 12. 12 • How to train this Generative Network? • One Solution - GAN • Generative - Generates new content • Adversarial Networks - Two networks with opposing objectives
- 13. 13 GAN • It consists of two deep networks: 1. Discriminator network 2. Generator network • The discriminator is supplied with both true and fake images and it tries to detect fake vs true images. Outputs a probability between 0 and 1 that the image is “true” • The generator tries to fool the discriminator into thinking that the fake images it generates are true images
- 14. Architecture of GAN x ~ Pdata(x) x ~ Pg(x) 14 Z ~ Pprior(z) (z;θg) (x;θd)
- 15. 15 Some terminology • Pdata(x) - Data distribution • Pg(x) - Generated distribution • Pprior(z) - Noise distribution • D(x;θd) - Discriminator function with parameters θ d • G(x;θg) - Generator function with parameters θ g
- 17. Discriminator Learning • Need to predict “1” for “true” images, and “0” for “fake” images • So the loss function (negative of log loss) for the discriminator is V’ = • We want to maximize this loss function. In other words, perform gradient ascent as θ d θ d + η Δ V’(θ d) 17
- 19. Generator Learning • Generator needs to fool the discriminator. • It needs the discriminator to output “1” for “fake” images. • So the loss function (log loss) for the generator is V’ = • We want to minimize this loss function. In other words, perform gradient descent as θ g θ g −η Δ V’(θ g) 19
- 20. Learning Algorithm 20 Train D K steps Train G 1 step
- 21. GAN – Learn a discriminator Discriminator v1 image 1/0 (real or fake) Random noise 1 1 1 1 0 0 0 0 Real images Sampled from DB: Update θ d v2 NN Generator v1 (FIX)
- 22. GAN – Learn a generator NN Generator v1 Random noise 0.13 Update the parameters of generator with BP Generator wants the output be classified as “real” (as close to 1 as possible) Use gradient descent to update the parameters in the generator, but fix the discriminator unchanged 1.0 v2 Train this Do not Train This They have Opposite objectives Discriminator v1 (FIX)
- 23. 23 • It has been shown in the paper that, if both the generator and discriminator are provided with enough capacity and training time we can get Pg(x) converge to Pdata(x)
- 24. 24 Takeaways • Generator models try to generate data from a given (complex) probability distribution • The generator tries to model the input data probability distribution Pg(x) = Pdata(x) • GAN uses adversarial method to train the Generator
- 25. 25 Applications • Generating images, videos, poems, some simple conversation • Text-to-image synthesis • Face Aging
- 26. Generating Anime Faces using GAN Reference: https://arxiv.org/pdf/1708.05509.pdf mages Source: Konachen Anime Faces (https://github.com/nagadomi/lbpcascade_animefa Training Samples: 50,000 pictures of 96x96 pixels
- 27. Generating Anime Faces using GAN 100 iterations
- 28. Generating Anime Faces using GAN 1000 iterations
- 29. Generating Anime Faces using GAN 2000 iterations
- 30. Generating Anime Faces using GAN 5000 iterations
- 31. Generating Anime Faces using GAN 10,000 iterations
- 32. Generating Anime Faces using GAN 20,000 iterations
- 33. Generating Anime Faces using GAN 50,000 iterations
- 34. Application Pose Guided Person Image Generation Reference: https://arxiv.org/pdf/1705.09368.pdf 34
- 35. Application CycleGAN Transform image from one domain (real) to another domain (Monet painting) 35 Reference: https://github.com/junyanz/CycleGAN
- 36. Application Text to image (StackGAN) Sketch Primitive shape and basic colours Generate high resolution with photo-realistic details Reference: https://arxiv.org/pdf/1612.03242v1.pdf 36
- 38. Application Face aging (Age-cGAN) Reference: https://arxiv.org/pdf/1702.01983.pdf 38
- 39. 39 References • https://towardsdatascience.com/understanding-generative-adversarial- networks-gans-cd6e4651a29 • https://arxiv.org/pdf/1406.2661.pdf • https://skymind.ai/wiki/generative-adversarial-network-gan • http://www.iangoodfellow.com/slides/2016-12-04-NIPS.pdf • http://slazebni.cs.illinois.edu/spring17/lec11_gan.pdf • https://cs.uwaterloo.ca/~mli/Deep-Learning-2017-Lecture7GAN.ppt • https://medium.com/@jonathan_hui/gan-some-cool-applications-of-gans- 4c9ecca35900
- 40. 40 Thank You!