SlideShare a Scribd company logo

Generative Adversarial Networks

M
Mustafa Yagmur

Generative adversarial networks (GANs) are a class of machine learning frameworks where two neural networks, a generator and discriminator, compete against each other. The generator learns to generate new data with the same statistics as the training set to fool the discriminator, while the discriminator learns to better distinguish real samples from generated samples. GANs have applications in image generation, image translation between domains, and image completion. Training GANs can be challenging due to issues like mode collapse.

Technology
Related topics:
Unsupervised Learning
1 of 42
Downloaded 52 times
1
2
Most read
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
Most read
28
29
30
31
32
33
34
35
36
37
38
39
Most read
40
41
42
GENERATIVE ADVERSARIAL NETWORKS by Mustafa Yagmur, Gregorio Nuevo Castro Dimitris Katsios, Suzana Ilic, Alisher Abdulkhaev, Basic Intuition and Applications
RoadMap Generative Models • Generative vs Discriminative Generative Adversarial Networks • Basic Framework • Loss • Training • Challenges Some Applications
Generative Models
Learn how to manipulate features Discriminative Generative Learn about a decision boundary between classes Figure Source : https://datawarrior.wordpress.com/2016/05/08/generative-discriminative-pairs/ classify, discriminate, differentiate, label generate new samples: new media, new music , new text Naïve Bayes , Hidden Markov models , Gaussian mixtures , Boltzman Machines , Variational Autoencoders , Generative Adversarial Networks Traditional Neural Networks , SVMs , Logistic regression ,Nearest neighbors , Random Forests
Generative Model Discriminative Model
Generative Model Discriminative Model
Make me a movie with.. Make me a video with.. Make me a picture with.. Make me a soundtrack with.. Make me a room with.. Make me a video game with.. % 100 machine generated life What is Exciting ! Image : techcrunch.com/Magic Pony Image : MIT Tech Review https://www.wired.com/story/ai-filmmaker-zone-out/
What does Generative Models do? generate a model distribution which close to true distribution model distribution target distribution target model target model Earlier attempt Latest attempt pdf pdf
Basic Generation Process Encoder Decoder Latent Features Neural Network Neural Network Do tricky operations on latent space VAE
What if a deep network used to judge the generation process and give feedback on generated images Novel Idea : Adversarial Training Since deep networks are very successful to recognize images like cat and dog, Generation Process External Judger
Generative Adversarial Networks
everything starts from a simple random noise.. Generative Model Random numbers
Z : Random Noise R : Real samples G : Generator D Discriminator Basic Framework Z R Switch Train Source : https://www.kdnuggets.com/2017/01/generative-adversarial-networks-hot-topic-machine-learning.html
G: Generator D: Discriminator R: Real Mutual Learning Bad counterfeit
Mutual Learning G: Generator D: Detective R: Real Better counterfeit
D(x) x More on the Framework Probability distribution of x Discriminator D(G(z)) G(z) z Probability distribution (logits) of G of z Generated fake image Generator Discriminator
D(G(z)) G(z) z More on the Framework Probability distribution (logits) of G of z Generated fake image Generator Discriminator D(x) Probability distribution of x Discriminator
Logits : Dx , Labels :1 Logits : Dg , Labels : 0 D_loss_real (Dx , ones) D_loss_fake (Dg , zeros) X : real images Z (Random Noise) generator discriminator D(x) tries to be near 1 LossES D_loss = Discriminator training Generator training D tries to make D(G(z)) near 0, Gz : fake images
Logits : Dx , Labels :1 Logits : Dg , Labels : 0 D_loss_real (Dx , ones) D_loss_fake (Dg , zeros) Gz : fake images X : real images Z (noise vector) generator discriminator D(x) tries to be near 1 LossES G tries to make D(G(z)) near 1 D_loss = G_loss = G_loss (Dg , ones) Logits : Dg , Labels : 1 Discriminator training Generator training D tries to make D(G(z)) near 0, Gz : fake images
Discriminator Generator Step 1. Train Discriminator Backpropagation Step 2. Train Generator Discriminator Generator Backpropagation only classifier mode Freeze generator Z Z (only generate samples) Update parameters Generate samples Freeze discriminator Update parameters
Discriminator Generator Step 1. Train Discriminator Backpropagation Step 2. Train Generator Discriminator Generator Backpropagation only classifier mode Freeze generator Z Z Only generate samples Update parameters Generate samples Freeze discriminator Update parameters
Challenges ! Mode collapse If one of them fails, the whole system fails discriminator much powerful generator too lenient discriminator too lenient Generator much powerful generating similar samples Generator explores some weakness in discriminator Top : desired case Bottom : collapsed case If else Metz , 2016 print
Types & Applications
You may usually find two main trends Research efforts on stabilizing GANs Investigating different use cases A Quick Overview
Github : hindupuravinash/the-gan-zoo The GAN Zoo Vanilla GAN aGAN bGAN cGAN dGAN … Check this out for the zoo !
Play Puzzle ! Change elements as you want and get new one :) Source Paper : https://arxiv.org/abs/1610.09585
Upsampling Downsampling Source Paper : https://arxiv.org/abs/1511.06434 Deep Convolutional GAN (DCGAN)
Use side information (classes) to guide learning process Conditional GAN (cGAN) Source Paper: https://arxiv.org/abs/1411.1784
Bidirectional GAN (BiGAN) Vanilla GAN BiGAN
Progressive GAN Source: Nvidia
Source Paper : https://arxiv.org/abs/1612.03242 Text to Photo-realistic Image Synthesis (StackGAN) Given a text description of a bird, generate a new bird image to match the text.
Source Paper : https://arxiv.org/pdf/1607.07539.pdf Image completion
Basic Arithmetic Operations (DCGAN) Source Paper : https://arxiv.org/abs/1511.06434
Image to Image Translation Source Paper : https://arxiv.org/abs/1611.07004
• Training difficulties • Importance of hyperparameter tuning • Counting, perspective issues • … Takeaways
Questions Let’s start with the most funny part ! and
References
[1] Generative Adversarial Networks Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, Yoshua Bengio https://arxiv.org/abs/1406.2661 [2] Understanding Generative Adversarial Networks Nooki Shibuya https://github.com/udacity/deep-learning/tree/master/gan_mnist [3] Generative Adversarial Networks Youtube Series Siraj Raval https://www.youtube.com/watch?v=0VPQHbMvGzg&t=1752s [4] How to Train a GAN? Tips and tricks to make GANs work Facebook AI Research: Soumith Chintala, Emily Denton, Martin Arjovsky, Michael Mathieu https://github.com/soumith/ganhacks [5] Generative Models Agustinus Kristiadi https://github.com/wiseodd/generative-models [6] Generative Models Analytics Vidhya , Introductory guide to Generative Adversarial Networks (GANs)
[7] Image Completion with Deep Learning in TensorFlow Brandon Amos http://bamos.github.io/2016/08/09/deep-completion/ [8] GAN MNIST Example in TensorFlow Udacity https://github.com/udacity/deep-learning/tree/master/gan_mnist [9] Generative Adversarial Networks Robin Richard http://www.rricard.me/machine/learning/generative/adversarial/networks/keras/tensorflow/2017/04/05/gans- part1.html [10] GAN Tutorial Series MLNotebook https://mlnotebook.github.io/post/GAN1/ [11] GAN Tutorial DL4J https://deeplearning4j.org/generative-adversarial-network [12] Generative Adversarial Networks Yunjey Choi https://www.slideshare.net/YunjeyChoi/generative-adversarial-networks-75916964
[13] Generative Adversarial Networks JaeJun Yoo https://www.slideshare.net/thinkingfactory/pr12-intro-to-gans-jaejun-yoo [14] Variational Autoencoders Ali Ghodsi https://www.youtube.com/watch?v=uaaqyVS9-rM&t=574s [15] Generative Models Fei Fei Li , Justin Johnson, Serena Yeung https://www.youtube.com/watch?v=5WoItGTWV54 [16] Machine Learning is Fun Part 7:GANs Adam Geitgey https://medium.com/@ageitgey/abusing-generative-adversarial-networks-to-make-8-bit-pixel-art-e45d9b96cee7 [17] Generative Adversarial Networks Lara Hulbert https://medium.com/@lehulbert/generative-adversarial-networks-85da3ec78fd2 [18] Upscaling images CSI-style with generative adversarial neural networks Geoffrey Litt http://geoffreylitt.com/2017/06/04/enhance-upscaling-images-with-generative-adversarial-neural-networks.html
Auxiliary Examples Counting , Perspective Problem
and Training Discriminator Generator Repeat for each epoch

More Related Content

PPTX
Generative Adversarial Networks (GAN)
Manohar Mukku
 
PDF
Introduction to Generative Adversarial Networks (GANs)
Appsilon Data Science
 
PPTX
A (Very) Gentle Introduction to Generative Adversarial Networks (a.k.a GANs)
Thomas da Silva Paula
 
PDF
GANs and Applications
Hoang Nguyen
 
PPTX
Generative Adversarial Networks (GANs)
Amol Patil
 
PDF
Deep Learning for Computer Vision: Generative models and adversarial training...
Universitat Politècnica de Catalunya
 
PDF
Generative adversarial networks
Yunjey Choi
 
PPTX
GANs Presentation.pptx
MAHMOUD729246
 
Generative Adversarial Networks (GAN)
Manohar Mukku
 
Introduction to Generative Adversarial Networks (GANs)
Appsilon Data Science
 
A (Very) Gentle Introduction to Generative Adversarial Networks (a.k.a GANs)
Thomas da Silva Paula
 
GANs and Applications
Hoang Nguyen
 
Generative Adversarial Networks (GANs)
Amol Patil
 
Deep Learning for Computer Vision: Generative models and adversarial training...
Universitat Politècnica de Catalunya
 
Generative adversarial networks
Yunjey Choi
 
GANs Presentation.pptx
MAHMOUD729246
 

What's hot (20)

PDF
Generative Adversarial Networks (GANs) - Ian Goodfellow, OpenAI
WithTheBest
 
PDF
Generative adversarial networks
남주 김
 
PDF
Basic Generative Adversarial Networks
Dong Heon Cho
 
PDF
GAN - Theory and Applications
Emanuele Ghelfi
 
PDF
Generative adversarial text to image synthesis
Universitat Politècnica de Catalunya
 
PDF
Introduction to Generative Adversarial Networks
BennoG1
 
PPTX
Generative adversarial networks
Ding Li
 
PDF
Generative Adversarial Networks and Their Applications
Artifacia
 
PPTX
Generative Adversarial Network (GAN)
Prakhar Rastogi
 
PDF
1시간만에 GAN(Generative Adversarial Network) 완전 정복하기
NAVER Engineering
 
PDF
A Short Introduction to Generative Adversarial Networks
Jong Wook Kim
 
PDF
Image-to-Image Translation with Conditional Adversarial Nets (UPC Reading Group)
Universitat Politècnica de Catalunya
 
PDF
Unsupervised learning represenation with DCGAN
Shyam Krishna Khadka
 
PPTX
Diffusion models beat gans on image synthesis
BeerenSahu
 
PDF
Deep Learning - Convolutional Neural Networks
Christian Perone
 
PDF
Variational Autoencoders For Image Generation
Jason Anderson
 
PDF
PR-409: Denoising Diffusion Probabilistic Models
Hyeongmin Lee
 
PDF
Convolutional Neural Network Models - Deep Learning
Mohamed Loey
 
PDF
Deep Generative Models
Chia-Wen Cheng
 
PDF
Introduction to Deep learning
Massimiliano Ruocco
 
Generative Adversarial Networks (GANs) - Ian Goodfellow, OpenAI
WithTheBest
 
Generative adversarial networks
남주 김
 
Basic Generative Adversarial Networks
Dong Heon Cho
 
GAN - Theory and Applications
Emanuele Ghelfi
 
Generative adversarial text to image synthesis
Universitat Politècnica de Catalunya
 
Introduction to Generative Adversarial Networks
BennoG1
 
Generative adversarial networks
Ding Li
 
Generative Adversarial Networks and Their Applications
Artifacia
 
Generative Adversarial Network (GAN)
Prakhar Rastogi
 
1시간만에 GAN(Generative Adversarial Network) 완전 정복하기
NAVER Engineering
 
A Short Introduction to Generative Adversarial Networks
Jong Wook Kim
 
Image-to-Image Translation with Conditional Adversarial Nets (UPC Reading Group)
Universitat Politècnica de Catalunya
 
Unsupervised learning represenation with DCGAN
Shyam Krishna Khadka
 
Diffusion models beat gans on image synthesis
BeerenSahu
 
Deep Learning - Convolutional Neural Networks
Christian Perone
 
Variational Autoencoders For Image Generation
Jason Anderson
 
PR-409: Denoising Diffusion Probabilistic Models
Hyeongmin Lee
 
Convolutional Neural Network Models - Deep Learning
Mohamed Loey
 
Deep Generative Models
Chia-Wen Cheng
 
Introduction to Deep learning
Massimiliano Ruocco
 
Ad

Similar to Generative Adversarial Networks (20)

PDF
M4L19 Generative Models - Slides v 3.pdf
yireme8491
 
PDF
Deep Generative Modelling (updated)
Petko Nikolov
 
PDF
Generative Adversarial Networks GAN - Xavier Giro - UPC TelecomBCN Barcelona ...
Universitat Politècnica de Catalunya
 
PPTX
GAN_SANTHOSH KUMAR_Architecture_in_network
SanthoshKumar229321
 
PDF
InfoGAN and Generative Adversarial Networks
Zak Jost
 
PDF
Deep Generative Models II (DLAI D10L1 2017 UPC Deep Learning for Artificial I...
Universitat Politècnica de Catalunya
 
PPTX
GAN Deep Learning Approaches to Image Processing Applications (1).pptx
RMDAcademicCoordinat
 
PPTX
Gans - Generative Adversarial Nets
SajalRastogi8
 
PDF
gan.pdf
Dr.rukmani Devi
 
PDF
Generative adversarial network_Ayadi_Alaeddine
Deep Learning Italia
 
PPTX
ICASSP 2018 Tutorial: Generative Adversarial Network and its Applications to ...
宏毅 李
 
PDF
Generative Adversarial Networks GAN - Santiago Pascual - UPC Barcelona 2018
Universitat Politècnica de Catalunya
 
PDF
Deep Generative Modelling
Petko Nikolov
 
PDF
Deep Generative Learning for All
Universitat Politècnica de Catalunya
 
PDF
Generative Models and Adversarial Training (D2L3 Insight@DCU Machine Learning...
Universitat Politècnica de Catalunya
 
PDF
What is a GAN Generative Adversarial Networks Guide.pdf
imoliviabennett
 
PPTX
gan-190318135433 (1).pptx
kiran814572
 
PPTX
GAN Generative Adversarial Networks.pptx
ssuser2624f71
 
PDF
Deep Generative Models - Kevin McGuinness - UPC Barcelona 2018
Universitat Politècnica de Catalunya
 
PPTX
GANs Deep Learning Summer School
Rubens Zimbres, PhD
 
M4L19 Generative Models - Slides v 3.pdf
yireme8491
 
Deep Generative Modelling (updated)
Petko Nikolov
 
Generative Adversarial Networks GAN - Xavier Giro - UPC TelecomBCN Barcelona ...
Universitat Politècnica de Catalunya
 
GAN_SANTHOSH KUMAR_Architecture_in_network
SanthoshKumar229321
 
InfoGAN and Generative Adversarial Networks
Zak Jost
 
Deep Generative Models II (DLAI D10L1 2017 UPC Deep Learning for Artificial I...
Universitat Politècnica de Catalunya
 
GAN Deep Learning Approaches to Image Processing Applications (1).pptx
RMDAcademicCoordinat
 
Gans - Generative Adversarial Nets
SajalRastogi8
 
gan.pdf
Dr.rukmani Devi
 
Generative adversarial network_Ayadi_Alaeddine
Deep Learning Italia
 
ICASSP 2018 Tutorial: Generative Adversarial Network and its Applications to ...
宏毅 李
 
Generative Adversarial Networks GAN - Santiago Pascual - UPC Barcelona 2018
Universitat Politècnica de Catalunya
 
Deep Generative Modelling
Petko Nikolov
 
Deep Generative Learning for All
Universitat Politècnica de Catalunya
 
Generative Models and Adversarial Training (D2L3 Insight@DCU Machine Learning...
Universitat Politècnica de Catalunya
 
What is a GAN Generative Adversarial Networks Guide.pdf
imoliviabennett
 
gan-190318135433 (1).pptx
kiran814572
 
GAN Generative Adversarial Networks.pptx
ssuser2624f71
 
Deep Generative Models - Kevin McGuinness - UPC Barcelona 2018
Universitat Politècnica de Catalunya
 
GANs Deep Learning Summer School
Rubens Zimbres, PhD
 
Ad

Recently uploaded (20)

PDF
Peak of Data & AI Encore- AI for Metadata and Smarter Workflows
Safe Software
 
PDF
Modernizing your data center with Dell and AMD
Principled Technologies
 
PDF
Machine learning based COVID-19 study performance prediction
IAESIJAI
 
PPTX
PA Analog/Digital System: The Backbone of Modern Surveillance and Communication
AVTRON Technologies LLC
 
PDF
Network Security Unit 5.pdf for BCA BBA.
Serpent6
 
PDF
cuic standard and advanced reporting.pdf
SimoneTitaniumTomase
 
PDF
KodekX | Application Modernization Development
KodekX
 
PDF
Empathic Computing: Creating Shared Understanding
Mark Billinghurst
 
PDF
Diabetes mellitus diagnosis method based random forest with bat algorithm
IAESIJAI
 
PPTX
Cloud computing and distributed systems.
kkkkkhan
 
PDF
NewMind AI Weekly Chronicles - August'25 Week I
NewMind AI
 
PDF
Chapter 3 Spatial Domain Image Processing.pdf
Getnet Tigabie Askale -(GM)
 
PPT
“AI and Expert System Decision Support & Business Intelligence Systems”
ZubinRadhakrishnan
 
PPTX
20250228 LYD VKU AI Blended-Learning.pptx
DatVoNgoc
 
PDF
Shreyas Phanse Resume: Experienced Backend Engineer | Java • Spring Boot • Ka...
SHREYAS PHANSE
 
PPT
Teaching material agriculture food technology
LiaRayya
 
PPTX
breach-and-attack-simulation-cybersecurity-india-chennai-defenderrabbit-2025....
defencerabbit Team
 
PDF
CIFDAQ's Market Insight: SEC Turns Pro Crypto
CIFDAQ
 
PDF
solutions_manual_-_materials___processing_in_manufacturing__demargo_.pdf
AbdullahSani29
 
PDF
GamePlan Trading System Review: Professional Trader's Honest Take
SOFTTECHHUB
 
Peak of Data & AI Encore- AI for Metadata and Smarter Workflows
Safe Software
 
Modernizing your data center with Dell and AMD
Principled Technologies
 
Machine learning based COVID-19 study performance prediction
IAESIJAI
 
PA Analog/Digital System: The Backbone of Modern Surveillance and Communication
AVTRON Technologies LLC
 
Network Security Unit 5.pdf for BCA BBA.
Serpent6
 
cuic standard and advanced reporting.pdf
SimoneTitaniumTomase
 
KodekX | Application Modernization Development
KodekX
 
Empathic Computing: Creating Shared Understanding
Mark Billinghurst
 
Diabetes mellitus diagnosis method based random forest with bat algorithm
IAESIJAI
 
Cloud computing and distributed systems.
kkkkkhan
 
NewMind AI Weekly Chronicles - August'25 Week I
NewMind AI
 
Chapter 3 Spatial Domain Image Processing.pdf
Getnet Tigabie Askale -(GM)
 
“AI and Expert System Decision Support & Business Intelligence Systems”
ZubinRadhakrishnan
 
20250228 LYD VKU AI Blended-Learning.pptx
DatVoNgoc
 
Shreyas Phanse Resume: Experienced Backend Engineer | Java • Spring Boot • Ka...
SHREYAS PHANSE
 
Teaching material agriculture food technology
LiaRayya
 
breach-and-attack-simulation-cybersecurity-india-chennai-defenderrabbit-2025....
defencerabbit Team
 
CIFDAQ's Market Insight: SEC Turns Pro Crypto
CIFDAQ
 
solutions_manual_-_materials___processing_in_manufacturing__demargo_.pdf
AbdullahSani29
 
GamePlan Trading System Review: Professional Trader's Honest Take
SOFTTECHHUB
 

Generative Adversarial Networks

  • 1. GENERATIVE ADVERSARIAL NETWORKS by Mustafa Yagmur, Gregorio Nuevo Castro Dimitris Katsios, Suzana Ilic, Alisher Abdulkhaev, Basic Intuition and Applications
  • 2. RoadMap Generative Models • Generative vs Discriminative Generative Adversarial Networks • Basic Framework • Loss • Training • Challenges Some Applications
  • 4. Learn how to manipulate features Discriminative Generative Learn about a decision boundary between classes Figure Source : https://datawarrior.wordpress.com/2016/05/08/generative-discriminative-pairs/ classify, discriminate, differentiate, label generate new samples: new media, new music , new text Naïve Bayes , Hidden Markov models , Gaussian mixtures , Boltzman Machines , Variational Autoencoders , Generative Adversarial Networks Traditional Neural Networks , SVMs , Logistic regression ,Nearest neighbors , Random Forests
  • 7. Make me a movie with.. Make me a video with.. Make me a picture with.. Make me a soundtrack with.. Make me a room with.. Make me a video game with.. % 100 machine generated life What is Exciting ! Image : techcrunch.com/Magic Pony Image : MIT Tech Review https://www.wired.com/story/ai-filmmaker-zone-out/
  • 8. What does Generative Models do? generate a model distribution which close to true distribution model distribution target distribution target model target model Earlier attempt Latest attempt pdf pdf
  • 9. Basic Generation Process Encoder Decoder Latent Features Neural Network Neural Network Do tricky operations on latent space VAE
  • 10. What if a deep network used to judge the generation process and give feedback on generated images Novel Idea : Adversarial Training Since deep networks are very successful to recognize images like cat and dog, Generation Process External Judger
  • 12. everything starts from a simple random noise.. Generative Model Random numbers
  • 13. Z : Random Noise R : Real samples G : Generator D Discriminator Basic Framework Z R Switch Train Source : https://www.kdnuggets.com/2017/01/generative-adversarial-networks-hot-topic-machine-learning.html
  • 14. G: Generator D: Discriminator R: Real Mutual Learning Bad counterfeit
  • 15. Mutual Learning G: Generator D: Detective R: Real Better counterfeit
  • 16. D(x) x More on the Framework Probability distribution of x Discriminator D(G(z)) G(z) z Probability distribution (logits) of G of z Generated fake image Generator Discriminator
  • 17. D(G(z)) G(z) z More on the Framework Probability distribution (logits) of G of z Generated fake image Generator Discriminator D(x) Probability distribution of x Discriminator
  • 18. Logits : Dx , Labels :1 Logits : Dg , Labels : 0 D_loss_real (Dx , ones) D_loss_fake (Dg , zeros) X : real images Z (Random Noise) generator discriminator D(x) tries to be near 1 LossES D_loss = Discriminator training Generator training D tries to make D(G(z)) near 0, Gz : fake images
  • 19. Logits : Dx , Labels :1 Logits : Dg , Labels : 0 D_loss_real (Dx , ones) D_loss_fake (Dg , zeros) Gz : fake images X : real images Z (noise vector) generator discriminator D(x) tries to be near 1 LossES G tries to make D(G(z)) near 1 D_loss = G_loss = G_loss (Dg , ones) Logits : Dg , Labels : 1 Discriminator training Generator training D tries to make D(G(z)) near 0, Gz : fake images
  • 20. Discriminator Generator Step 1. Train Discriminator Backpropagation Step 2. Train Generator Discriminator Generator Backpropagation only classifier mode Freeze generator Z Z (only generate samples) Update parameters Generate samples Freeze discriminator Update parameters
  • 21. Discriminator Generator Step 1. Train Discriminator Backpropagation Step 2. Train Generator Discriminator Generator Backpropagation only classifier mode Freeze generator Z Z Only generate samples Update parameters Generate samples Freeze discriminator Update parameters
  • 22. Challenges ! Mode collapse If one of them fails, the whole system fails discriminator much powerful generator too lenient discriminator too lenient Generator much powerful generating similar samples Generator explores some weakness in discriminator Top : desired case Bottom : collapsed case If else Metz , 2016 print
  • 24. You may usually find two main trends Research efforts on stabilizing GANs Investigating different use cases A Quick Overview
  • 25. Github : hindupuravinash/the-gan-zoo The GAN Zoo Vanilla GAN aGAN bGAN cGAN dGAN … Check this out for the zoo !
  • 26. Play Puzzle ! Change elements as you want and get new one :) Source Paper : https://arxiv.org/abs/1610.09585
  • 27. Upsampling Downsampling Source Paper : https://arxiv.org/abs/1511.06434 Deep Convolutional GAN (DCGAN)
  • 28. Use side information (classes) to guide learning process Conditional GAN (cGAN) Source Paper: https://arxiv.org/abs/1411.1784
  • 31. Source Paper : https://arxiv.org/abs/1612.03242 Text to Photo-realistic Image Synthesis (StackGAN) Given a text description of a bird, generate a new bird image to match the text.
  • 32. Source Paper : https://arxiv.org/pdf/1607.07539.pdf Image completion
  • 33. Basic Arithmetic Operations (DCGAN) Source Paper : https://arxiv.org/abs/1511.06434
  • 34. Image to Image Translation Source Paper : https://arxiv.org/abs/1611.07004
  • 35. • Training difficulties • Importance of hyperparameter tuning • Counting, perspective issues • … Takeaways
  • 36. Questions Let’s start with the most funny part ! and
  • 38. [1] Generative Adversarial Networks Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, Yoshua Bengio https://arxiv.org/abs/1406.2661 [2] Understanding Generative Adversarial Networks Nooki Shibuya https://github.com/udacity/deep-learning/tree/master/gan_mnist [3] Generative Adversarial Networks Youtube Series Siraj Raval https://www.youtube.com/watch?v=0VPQHbMvGzg&t=1752s [4] How to Train a GAN? Tips and tricks to make GANs work Facebook AI Research: Soumith Chintala, Emily Denton, Martin Arjovsky, Michael Mathieu https://github.com/soumith/ganhacks [5] Generative Models Agustinus Kristiadi https://github.com/wiseodd/generative-models [6] Generative Models Analytics Vidhya , Introductory guide to Generative Adversarial Networks (GANs)
  • 39. [7] Image Completion with Deep Learning in TensorFlow Brandon Amos http://bamos.github.io/2016/08/09/deep-completion/ [8] GAN MNIST Example in TensorFlow Udacity https://github.com/udacity/deep-learning/tree/master/gan_mnist [9] Generative Adversarial Networks Robin Richard http://www.rricard.me/machine/learning/generative/adversarial/networks/keras/tensorflow/2017/04/05/gans- part1.html [10] GAN Tutorial Series MLNotebook https://mlnotebook.github.io/post/GAN1/ [11] GAN Tutorial DL4J https://deeplearning4j.org/generative-adversarial-network [12] Generative Adversarial Networks Yunjey Choi https://www.slideshare.net/YunjeyChoi/generative-adversarial-networks-75916964
  • 40. [13] Generative Adversarial Networks JaeJun Yoo https://www.slideshare.net/thinkingfactory/pr12-intro-to-gans-jaejun-yoo [14] Variational Autoencoders Ali Ghodsi https://www.youtube.com/watch?v=uaaqyVS9-rM&t=574s [15] Generative Models Fei Fei Li , Justin Johnson, Serena Yeung https://www.youtube.com/watch?v=5WoItGTWV54 [16] Machine Learning is Fun Part 7:GANs Adam Geitgey https://medium.com/@ageitgey/abusing-generative-adversarial-networks-to-make-8-bit-pixel-art-e45d9b96cee7 [17] Generative Adversarial Networks Lara Hulbert https://medium.com/@lehulbert/generative-adversarial-networks-85da3ec78fd2 [18] Upscaling images CSI-style with generative adversarial neural networks Geoffrey Litt http://geoffreylitt.com/2017/06/04/enhance-upscaling-images-with-generative-adversarial-neural-networks.html
  • 41. Auxiliary Examples Counting , Perspective Problem