SlideShare uses cookies to improve functionality and performance, and to provide you with relevant advertising. If you continue browsing the site, you agree to the use of cookies on this website. See our User Agreement and Privacy Policy.
SlideShare uses cookies to improve functionality and performance, and to provide you with relevant advertising. If you continue browsing the site, you agree to the use of cookies on this website. See our Privacy Policy and User Agreement for details.
Successfully reported this slideshow.
Activate your 14 day free trial to unlock unlimited reading.
GAN Lab: Understanding Complex Deep Generative Models using Interactive Visual Experimentation
Minsuk Kahng, Nikhil Thorat, Duen Horng (Polo) Chau, Fernanda Viégas, Martin Wattenberg.
GAN Lab: Understanding Complex Deep Generative Models using Interactive Visual Experimentation.
IEEE Transactions on Visualization and Computer Graphics, 25(1) (VAST 2018).
Minsuk Kahng, Nikhil Thorat, Duen Horng (Polo) Chau, Fernanda Viégas, Martin Wattenberg.
GAN Lab: Understanding Complex Deep Generative Models using Interactive Visual Experimentation.
IEEE Transactions on Visualization and Computer Graphics, 25(1) (VAST 2018).
GAN Lab: Understanding Complex Deep Generative Models using Interactive Visual Experimentation
1.
GAN Lab
Understanding Complex Deep Generative Models
using Interactive Visual Experimentation
Georgia Tech Google
Minsuk
Kahng
Nikhil
Thorat
Polo
Chau
Fernanda
Viégas
Martin
Wattenberg
Georgia Tech Google Google
PAIR | People + AI Research Initiative
2.
2
Deep learning visualization tools presented at VIS
Most tools are designed for experts
3.
Many non-experts want to learn ML
3
Chris Olah’s Blog Andrej Karpathy’s Demo
4.
Many non-experts want to learn ML
3
Chris Olah’s Blog Andrej Karpathy’s Demo
Can our VIS community help this population?
5.
5
TensorFlow Playground
TensorFlow Playground was a great success
http://playground.tensorflow.org
7.
Generative Adversarial Networks (GANs)
6
Hard to understand and train even for experts
“the most interesting idea in the last 10 years in ML”
- Yann LeCun
Face images generated by BEGAN [Berthelot et al., 2017]
8.
Why are GANs hard to understand?
Because a GAN uses two competing neural networks
7
Discriminator
spots fake
Police
spots fake bills
Generator
synthesizes outputs
Counterfeiter
makes fake bills
9.
7
How to explain this concept using visualization?
Discriminator
spots fake
Police
spots fake bills
Generator
synthesizes outputs
Counterfeiter
makes fake bills
10.
GAN Lab
First Interactive Tool for Learning GANs in Browser
11.
Key contributions of GAN Lab
• Novel visualization of GAN’s training process
• Interactive model training
• Browser-based implementation
8
12.
How did we visualize GANs?
9
Discriminator
spots fake
Police
spots fake bills
Generator
synthesizes outputs
Counterfeiter
makes fake bills
13.
2D data distribution, instead of high-dimensional images
10
What type of data to visualize?
Discriminator
(Police)
Generator
(Counterfeiter)
14.
2D data distribution, instead of high-dimensional images
1411
What type of data to visualize?
Discriminator
(Police)
Generator
(Counterfeiter)
1. Easier to visualize data distribution
2. Easier for learners to track dynamics
Why 2D data points?
16.
How to visually explain the generator?
13
Generator
(Counterfeiter)
17.
How to visually explain the generator?
map an input point
into a new position
random
14
Generator
(Counterfeiter)
18.
How to visually explain the generator?
map an input point
into a new position
random
Manifold
?
14
Generator
(Counterfeiter)
19.
Mixture of
two Gaussians
15
How to visually explain the generator?
20.
How to visualize the discriminator?
16
Generator
(Counterfeiter)
Discriminator
(Police)
21.
2D heatmap, to represent its binary classification
17
How to visualize the discriminator?
Samples in this region are
likely real.
Samples are likely fake.
24.
1. Building mental
models for GANs
How does it help?
25.
1. Building mental
models for GANs
2. Tracking data flow
How does it help?
26.
1. Building mental
models for GANs
2. Tracking data flow
3. Locating
hyperparameters
How does it help?
27.
GAN Lab broadens education access
24
Conventional Deep Learning Visualization
in JavaScript
in Python with GPU
Model Training
Visualization
$$$
28.
Everything done in browser, powered by TensorFlow.js
GAN Lab broadens education access
25
Accelerated by WebGL
in JavaScript
Visualization
also in JavaScript
Model Training
29.
GAN Lab is Live!
20K visitors, 100+ countries 1.9K Likes 800+ Retweets
Try at bit.ly/gan-lab
30.
Georgia Tech
Google PAIR
Minsuk Kahng
Nikhil Thorat
Polo Chau
Fernanda Viégas
Martin Wattenberg
Georgia Tech
Google PAIR
Google PAIR
minsuk.comLearning and Playing with
GANs in your browser!
PAIR People + AI Research|
GAN Lab
|
Try at bit.ly/gan-lab