Generative Adversarial Networks

Deck from the talk given at 'AI DAY at Treebo', on 30th June 2018. Introduces one to the concept as well as the latest research trends in the field of GAN.
This deck features:
1. Motivation behind GAN concept.
2. Explaining the Adversarial process as well as common applications.
3. 2 New GAN architectures: Triple and Cycle GAN.

Transcript

1. Generative Adversarial Networks (GAN) By Mahir Jain
2. About Me I’m a student at National Institute of Technology, Karnataka, Surathkal (NITK). I am pursuing a B.Tech in Computer Science and Engineering. Interests: ML, DL, Cryptography and Computer Architecture.
3. GAN - A Background • Paper was presented at 2016 NIPS proceedings by Ian Goodfellow. Published in 2014. • “A framework for estimating generative models via an adversarial process.” • Generative models: Those that can sample a given distribution of data. • GAN’s are an unsupervised learning technique.
4. Minor Detour: The use of Generative Models • Simulate possible future states of a RL environment. • Very useful when one suffers from a lack of data. • Useful in Semi-Supervised Learning applications - Where there is negligible labelled data. • Another application - Image to Image Translation.
5. Example of Image to Image Translation Turn Sketches into actual images! Essentially, it is the inverse of the edge detection process that majority of the existing Deep Learning models use.
6. GAN - The Applications • Input: Low resolution image , Output: High resolution image. ◦ Why does this work using GAN’s ? - Multimodal Outputs. • Convert a text description to an image. • Sampling new drugs to cure diseases.
7. GAN over other Generative Models • Other existing models include ◦ Change of Variable Models: Latent space size can’t be varied. ◦ WaveNet: Very Slow despite good results. ◦ Variational Autoencoder: Poor quality results. High risk involved. • GAN’s overcome these issues to a large extent, although a quantitative measurement is hard to obtain.
8. GAN’s: The Adversarial Game • Train 2 adversaries: A Generator(G) and a Discriminator(D), that are both differentiable functions, typically Artificial Neural Networks (ANN’s), using simple Backpropagation. • The samples are estimated using simple forward propagation. • Discriminator needed only for the training step. Once results are satisfactory, we can use the trained Generator to sample the given distribution.
9. z ~ p z G x fake data x real D Fake Real
10. Revisiting Curing Diseases Now that we know the GAN fundamentals - How can we apply this to solve a real world problem?
11. My own GAN Experiment on the famous MNIST Dataset. • I implemented a simple feed-forward Neural Network for both G and D, on the famous MNIST Dataset. • Despite hardware limitations, that significantly reduced training speed, the results were decent. • Scope for good results is immense with good resources (High end GPU’s will significantly speed up computation.)
12. I WENT FROM HERE to HERE
13. Taking GAN to the next level….
14. Introducing the TRIPLE GAN • Paper published in 2017. • Traditional GAN faces 2 problems: ◦ Ideal/Saddle/Equilibrium point too hard to reach. ◦ Semantics (Data labels) Ignored. • Triple GAN proposes a way to overcome these issues.
15. • New player in the game: Classifier (C) • Before: Just generation of similar data(Eg: images) sufficient. • Now: Want more organised, meaningful output (Label the output), so we generate data-label pairs, instead of just data. • In short, we wish to CLASSIFY and GENERATE. • Useful for Semi-Supervised Learning (SSL)
16. Classifier Generator Discriminator Pseudo Labels Pseudo Data Real Data Real Labels Accept/Accept only one of the 2/Reject Both
17. MY TRIPLE GAN EXPERIENCE • CIFAR10 Dataset • Challenging task! • Technology/Framework Used : Google’s TensorFlow framework.
18. TRIPLE GAN: LESSONS LEARNT
19. Introducing the CYCLE GAN • Paper published in 2017. • Optimised to solve : Unpaired Image to Image Translation.
21. Why is this technique important? • Traditionally, training requires pair of images, ie, source and target both. • Unpaired Image to Image translation is able to work with just source images. The input is not pairs of images, rather it is
22. WHERE IS THE CYCLE? • If X -> Y, the model must also verify whether Y -> X. • Analogy: If we translate from English to French, we must be able to do so from French to English as well. (Transitivity)
24. In Conclusion….
25. Thank You!