Generative Adversarial
Networks (A.K.A GANs)
“Building something from nothing”
A (very) friendly introduction to
Csongor Barabasi
Artificial Intelligence Meetup Cluj Napoca
Who am I
Csongor Barabasi
Founder of Deep Learning School by Csongi
AI Student @ University of Birmingham
So … why GANs?
“Generative Adversarial Networks (GAN’s) are the
coolest thing since sliced bread. They are the coolest
thing in the last 10 years in ML”
Yann LeCun
Director of AI Research
Facebook
Machine Learning
“Machine Learning is the subfield of artificial intelligence
and it is based on algorithms that can learn from data
without relying on rule-based programming.”
McKinsey & Co.
Artificial
Intelligence
Machine
Learning Deep
Learning
Supervised Learning Unsupervised Learning
Supervised Learning
Size (sq.ft) Nr. bedrooms Price $
3100 7 60000
500 2 20000
2000 4 45000
Unsupervised Learning
Data
+ =
Keep feeding lots of data into your neural network until it yields a working model.
• Good for applications using existing, large, labelled datasets
• … but not suitable for niche applications, with small dataset
• Diagnosing a rare medical disease
Examples: FaceScrub (107,818), Skin Segmentation Dataset (245,057),
MNIST (60000), CIFAR-10/100 (60000)
What do we want to achieve
with AI?
1. Automated systems
• Not too smart
Artificial General Intelligence = handling novel tasks without explicit training
2. ML / DL
• The smartest we can get at the moment
3. GANs
• Less data, more intelligence
Convolutional Neural Networks
Convolutional Neural Networks
Inference vs Training
GAN Example
GAN Example
Go to the party
Get in or get
rejected
Feedback to make the ticket better
GAN
Random
Noise
Generator
“The Forger”
Synthesized
Data
Real
Data
Switch Discriminator
“The security”
Real
Fake
Feedback
Generative Model
• Tries to learn a representation of the data
• Then tries to create generate something new based on the
representation learned
• Composed of Deconvolution layers
Discriminative Model
• Learns a function that maps an input X to an output y
• Does classification
• GAN is an unsupervised technique trained in a supervised manner
When does the training end?
“It should happen until the generator exactly reproduces
the true data distribution and the discriminator is
guessing at random, unable to find a difference.”
Open AI
Cool thing #1
Data Synthesis
• Generate more data (usually for Deep Learning)
Cool thing #2
Learn translation between
datasets
• Low resolution to high resolution images
Cool thing #3
Adversarial attacks on DL
systems
• Scares the hell out of Deep Learning people
Interesting GAN applications
Image Generation
Text to Image
Neural Style Transfer
Interactive Image Generation
https://www.youtube.com/watch?v=5jfViPdYLic
https://www.youtube.com/watch?v=HOn8437TWPA
Sketch to Image
Anime Character Generation
3D Object Generation
Vector Arithmetic on images
Wrapping up
1. GANs help us apply Deep Learning to new niches
• Humans do not have to create huge, labeled datasets. GANs can do it
• Smarter, more autonomous learning
• A toolkit for manipulating and merging datasets
2. GANs are disrupting deep learning
• Scary GAN based adversarial attacks
• Humans do not have to create huge, labeled datasets. GANs can do it
3. GANs hint at new AI capabilities
• More compelling human-AI collaboration
Thank you! :D
“What I cannot create, I cannot understand.”
Richard Feynman

A friendly introduction to GANs