Week 6

Question 1

Common uses of GANs

Answer 1

Produce new content (eg extra digits for MNIST)
Text to image generation
Image to image translation
Increasing image resolution
Predicting next video frame

Question 2

Explicit model: trackable density

Answer 2

Question 3

Approximate density

Answer 3

Variational auto encoder

Question 4

Training objective function for VAE

Answer 4

L2 loss

Question 5

NN structure for auto encoder

Answer 5

Question 6

VAE show loss breakdown across encoding and decoding

Answer 6

Question 7

Problem that requires GANs and how it is a solution

Answer 7

Problem is we want to sample from complex and high dimensional training sample distribution

Solution is we sample from a simple distribution (eg random noise) and transforming it to training distribution (by generator network)

Question 8

GAN structure

Answer 8

Question 9

Data distributions for GANs

Answer 9

Question 10

Notation for GAN

Answer 10

Question 11

Loss function for GAN

Answer 11

Question 12

Formulate training for GAN

Answer 12

Question 13

Minibar y SGD for GAN

Answer 13

Question 14

GAN problem of non convergence

Answer 14

Question 15

GAN problem of diminished gradient

Answer 15

Question 16

Other problems for GANs

Answer 16

Mode collapse

Unbalance between generator and discriminator (overfitting, eg discriminator works too well)

Highly sensitive to hyper parameters

Question 17

Improving GANs with network design

Answer 17

Question 18

Improving GANs with cost functions

Answer 18

Question 19

Improving GANs with optimisation (experience replays)

Answer 19

Question 20

Improving GANs with optimisation - training with labels

Answer 20

Question 21

Improving GANs with optimisation - adding noise

Answer 21

Question 22

Improving GANs with optimisation - unrolling GAN

Answer 22

Question 23

Predict move of discriminator by unrolling

Answer 23