GANs

Question 1

Name the difference between explicit density estimation and implicit density estimation. Which of the two categories do GANs belong to?

Answer 1

explicit density estimation: explicitly define and solve the distribution from the data
implicit density estimation: implicitly learn model that can sample from the distribution of the data without explicitly difining it

GANs belongs to implicit desnity estimation.

Question 2

Write minmax objective functions.

Answer 2

minmax = [Ex~pdata log(Dθd(x) + Ez~pz log (1 - Dθd(G(z)))]

Question 3

Mention 2 advantages and 2 disadvantages of adversial training

Answer 3

Advantages: With a criterion for evaluating the quality of generated samples, the
gradient with respect to the network parameters can be computed; the parameters can be updated to generate samples with a
little better quality

Disadvantages:

Question 4

How to evaluate GANs

Answer 4

Sample quality: Does the image look real?; diverse sample: It generates one or many plausible samples? ;Generalization: It has a good generalization or simply memorized the training sample?

Question 5

GANs are evaluated using different metrics compared to typical metrics for discriminative model. Explain the reason

Answer 5

GANs generates new data samples instead of classifying existing data. Alternative metrics as Inceptions score or Frèchet evaluates diverse and sample quality from generated data.

Question 6

You trained a GAN. Inception score excellent, fréchet inception distance poor. What could be a reasonable reason for the observed behavior?

Answer 6

The inception score just evaluates sample quality and class diversity (if generates sample from all classes), but do not evaluate the sample diversity. The fréchet can evaluate the sample diversity using the fréchet distance. So in that case, the GANs should have a bad sample diversity

Question 7

3 central properties for good GANs; Name 3 methods to evaluate those

Answer 7

Sample quality: Does the image look real?; diverse sample: It generates one or many plausible samples? ;Generalization: It has a good generalization or simply memorized the training sample?

Sample quality: Inception score; Diverse sample: Frètche inception distance; Generalization: Nearest Neighbor

Question 8

Pix2Pix: Its usability is limited. What factor limits this method?

Answer 8

it requires pairs of
samples with pixel-to-pixel correspondence.

Question 9

How does Cycle GAN address the problem of Pix2Pix?

Answer 9

The cycle GAN architecture learns to do image-to-image translation from unpaired-data