Generative Adversarial Networks (GANs) Advanced Section #8

Advanced Section #8: Generative Adversarial Networks (GANs)

CS109B Data Science 2

Vincent Casser

Pavlos Protopapas

Outline

Concept and Math Applications Common Problems Wasserstein GANs, Conditional GANs and CycleGANs Troubleshooting GANs Hands-on: Building an Image GAN in Keras Influential Papers and References

CS109B, PROTOPAPAS, GLICKMAN

Concept

Generator Job: Fool discriminator

Discriminator Job: Catch lies of the generator

Real

Generated

"Both are pandas!"

Confidence: 0.9997

Confidence: 0.1617

"Nope"

CS109B, PROTOPAPAS, GLICKMAN

Concept

Generator Job: Fool discriminator

Discriminator Job: Catch lies of the generator

Generated

Real

"Both are pandas!"

Confidence: 0.3759

Confidence: 1.0

"Good try..."

CS109B, PROTOPAPAS, GLICKMAN

GAN Structure

Generator Job: Fool discriminator

Noise z

Sample G(z) G

Discriminator Job: Catch lies of the generator

Sample D

x (real)

G(z) (fake)

Score

D(x) -> 1 D(G(z)) -> 0

CS109B, PROTOPAPAS, GLICKMAN

Math in a nutshell

Generator

m: Number of samples z: Random noise samples x: Real samples

How realistic are the generated samples?

G wants to maximize this.

Discriminator

Make sure real samples are classified as being real.

Make sure generated samples are classified as unreal.

D wants to maximize this. D wants to minimize this.

CS109B, PROTOPAPAS, GLICKMAN

Math in a nutshell

Generator

m: Number of samples z: Random noise samples x: Real samples

Discriminator

x

G(z)

Generator Discriminator

D(x) = 0.9997

D(G(z)) = 0.1617

-

1.0

1.0

0.0

Targets

CS109B, PROTOPAPAS, GLICKMAN

Applications

(Conditional) synthesis

Font generation Text2Image 3D Object generation

Data augmentation

Aiming to reduce need for labeled data GAN is only used as a tool enhancing the training process of another model

Style transfer and manipulation

Face Aging Painting Pose estimation and manipulation Inpainting Blending

Signal super resolution

 ................
................

In order to avoid copyright disputes, this page is only a partial summary.

Google Online Preview   Download