Generative Adversarial Networks (GANs) at the Data Science Meetup Luxembourg (papers we like session)
0 likes241 views
Generative adversarial networks (GANs) are a class of machine learning frameworks where two neural networks compete against each other in a game framework. One network generates new data instances, while the other evaluates them for authenticity compared to real data. This adversarial process allows generating synthetic data that matches the distribution of real data. The summary discusses how GANs work, some applications like image generation, and challenges with training instability and evaluation.
![In contrast, generative models model the conditional distribution
P(data | Y = label)
and classify using Bayes theorem
we can obtain P(label | data).
E.g. in naive Bayes
P(Y = y| data) = P(Y = y) * [P(data_i | Y = y )]
Generative Adversarial Networks
9](https://image.slidesharecdn.com/datasciencemeetupluxembourgpaperswelikegans-180603083843/85/Generative-Adversarial-Networks-GANs-at-the-Data-Science-Meetup-Luxembourg-papers-we-like-session-9-320.jpg)
Generative Adversarial Networks (GANs) at the Data Science Meetup Luxembourg (papers we like session)
- 1. Papers We Like “Generative Adversarial Networks” by Ian Goodfellow (2014) May 24th, 2018 Data Science Luxembourg Meetup at the Interdisciplinary Centre for Security, Reliability and Trust Presented by Chris Hammerschmidt, PhD hammerschmidt@posteo.de
- 6. Discriminative models (such as SVMs) construct a decision boundary. Generative Adversarial Networks 6
- 7. discriminative models construct a decision boundary by modelling the conditional distribution P(label | data) Generative Adversarial Networks 7
- 8. In contrast, generative models model the conditional distribution P(data | Y = label) and classify using Bayes theorem we can obtain P(label | data). Generative Adversarial Networks 8
- 9. In contrast, generative models model the conditional distribution P(data | Y = label) and classify using Bayes theorem we can obtain P(label | data). E.g. in naive Bayes P(Y = y| data) = P(Y = y) * [P(data_i | Y = y )] Generative Adversarial Networks 9
- 10. A Generative Adversarial Situation 10
- 11. A Generative Adversarial Game 11 present fake or real note Feedback
- 12. ● A zero-sum game between two players 1) Discriminator D 2) Generator G ● D tries to discriminate between: ○ A sample from the data distribution p_data ○ A sample from the generator G ● G tries to trick D by generating samples that are hard for D to distinguish from data A Generative Adversarial Function Game 12
- 13. Training is done by taking turns between training generative model G and the discriminative model D. A Generative Adversarial Network Situation 13 Image credit: https://www.kdnuggets.com/2016/07/mnist-generative-adversarial-model-keras.html
- 14. The Architecture of a GAN 14 Differentiable function -> gradient descent Neural networks -> backprop Image credit: http://www.iangoodfellow.com/slides/
- 15. The GAN game 15 Image credit: http://www.iangoodfellow.com/slides/
- 16. Learning / Density Estimation 16 Image credit: http://www.iangoodfellow.com/slides/
- 17. Learning / Density Estimation 17 Image credit: http://www.iangoodfellow.com/slides/
- 18. Learning / Density Estimation 18 Image credit: http://www.iangoodfellow.com/slides/
- 19. Learning / Density Estimation 19 Image credit: http://www.iangoodfellow.com/slides/
- 20. A Generative Adversarial Network Situation 20 Image credit: https://www.slideshare.net/ckmarkohchang/generative-adversarial-networks
- 21. Does it work? 21
- 22. Context & Applications 22 Answer 1 Image credit: Ian Goodfellow’s original paper
- 23. Context & Applications 23 Answer 2 (i) Image credit: Ian Goodfellow’s original paper
- 24. Context & Applications 24 In the process, the global minimum minimizes the Jensen–Shannon divergence between the model’s distribution p_G and the data generating process p_data. Answer 2 (ii)
- 26. ● Theoretical properties: ○ Unique global optimum ○ Optimum corresponds to data distribution ○ Convergence is guaranteed ● BUT IN REALITY ○ Assumes infinite data ○ Assumes infinite capacity ○ Assumes correct updating of p_G Using SGD we have no proof of convergence! Assessment 26
- 27. Context & Applications and Why do we even care? 27
- 29. ● It seems to work pretty well for images and movies ● It avoids problems alternative generative approaches have: ○ No direct representation of p(x) needed ○ No MCMC or other sampling methods required ○ Everything is tractable computationally ○ No data goes to G directly, only gradients from D ● Unsupervised learning methods are useful! Why do we care? 29
- 30. Let’s look at some examples 30
- 31. DiscoGAN 31 https://arxiv.org/abs/1703.05192 Learning to Discover Cross-Domain Relations with Generative Adversarial Networks Taeksoo Kim, Moonsu Cha, Hyunsoo Kim, Jung Kwon Lee, Jiwon Kim
- 32. Radford’s Representation Learning Paper 32 https://arxiv.org/abs/1511.06434 Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks Alec Radford, Luke Metz, Soumith Chintala
- 33. Radford’s Representation Learning Paper 33
- 35. Isola’s Image-to-Image Translation 35 https://arxiv.org/abs/1611.07 004 Image-to-Image Translation with Conditional Adversarial Networks Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros
- 36. PassGAN 36 https://arxiv.org/abs/1709.00440 PassGAN: A Deep Learning Approach for Password Guessing Briland Hitaj, Paolo Gasti, Giuseppe Ateniese, Fernando Perez-Cruz
- 37. MalGAN 37 https://arxiv.org/abs/1702 .05983 Generating Adversarial Malware Examples for Black-Box Attacks Based on GAN Weiwei Hu, Ying Tan
- 38. Long List of Tasks in Data Science 38 Gaining insight into data in unsupervised settings is often the real difficulty in data science, not the fancy algorithms or “simple” supervised settings often considered in machine learning research with standardized problems and data sets: ● Lack of labels, missing values, confidential data ● We might not know what good labels are ● We might not know relationships between features ● Data is often only means to an end (simulation, policy learning)
- 39. Caveats 39
- 40. ● How to train GANs correctly? Training is notoriously unstable and suffers from some problems such as “mode collapse”, and many proposals to improve training exist. ● How to evaluate them effectively? They are trained on a density estimation task but often ultimately used for something else. Results shows that criteria such as log-likelihood are at odds with visual quality measures such as inception scores. There are open problems 40
- 42. Thanks! Questions? Do you have challenging data science problems? Contact me at hammerschmidt at posteo döt de Examples and inspiration from Ian Goodfellow’s slides at http://www.iangoodfellow.com/slides/ 42