Generative adversarial networks (GANs) quiz

Question 1

What is the minimax objective of the original GAN, and what does it represent at the global optimum?

Accepted Answer

At the optimum p_g = p_data and D(x) = 1/2 everywhere

Answer

At the optimum D(x) = 1 for all real and fake samples

Answer

At the optimum the generator outputs the mean of the data

Answer

At the optimum the discriminator loss is maximized to infinity

Question 2

For a fixed generator, what is the optimal discriminator D*(x) in the original GAN formulation?

Accepted Answer

D*(x) = p_data(x) / (p_data(x) + p_g(x))

Answer

D*(x) = p_g(x) / (p_data(x) + p_g(x))

Answer

D*(x) = p_data(x) / p_g(x)

Answer

D*(x) = 1/2 for all x

Question 3

Why is the non-saturating generator loss preferred over the original minimax generator loss?

Accepted Answer

It provides stronger gradients when samples are easily rejected, avoiding saturation

Answer

It guarantees convergence to a Nash equilibrium

Answer

It removes the need for a discriminator

Answer

It minimizes the exact JS divergence

Question 4

What is mode collapse in GANs and what is one common cause?

Accepted Answer

The generator produces limited diversity, covering only a few data modes

Answer

The discriminator always outputs 0.5

Answer

The generator perfectly reproduces all training images

Answer

The loss diverges to infinity

Question 5

How does the Wasserstein GAN (WGAN) loss differ from the original GAN, and what problem does it address?

Accepted Answer

It optimizes the Wasserstein-1 distance for smoother gradients under disjoint supports

Answer

It uses a sigmoid critic to bound the loss between 0 and 1

Answer

It removes the generator entirely

Answer

It maximizes KL divergence instead of JS

Question 6

In WGAN, how is the 1-Lipschitz constraint on the critic originally enforced, and why is it needed?

Accepted Answer

Weight clipping to a fixed range to satisfy Kantorovich-Rubinstein duality

Answer

Dropout on the critic

Answer

Batch normalization in the generator

Answer

Label smoothing on real samples

Question 7

What is the gradient penalty in WGAN-GP and how is it computed?

Accepted Answer

Penalize the critic's gradient norm toward 1 at interpolated points

Answer

Penalize the generator weights toward zero

Answer

Clip the discriminator output to [0,1]

Answer

Add Gaussian noise to the loss

Question 8

What architectural guidelines does DCGAN introduce for stable convolutional GAN training?

Accepted Answer

Strided/transposed convs, batchnorm, ReLU in G and LeakyReLU in D

Answer

Max pooling everywhere and sigmoid activations

Answer

Only fully connected layers with dropout

Answer

Recurrent layers in both networks

Question 9

How does a conditional GAN (cGAN) incorporate label information, and what does it enable?

Accepted Answer

By feeding labels to both G and D to model p(x|y)

Answer

By training a separate GAN per class only

Answer

By adding labels only to the loss, not the networks

Answer

By removing the discriminator

Question 10

What is the key innovation of CycleGAN that enables unpaired image-to-image translation?

Accepted Answer

Cycle-consistency loss enforcing reconstruction after round-trip translation

Answer

Paired supervised pixel loss

Answer

A single generator shared across domains

Answer

Using only the Wasserstein distance

Question 11

What problem does Progressive Growing of GANs (ProGAN) solve and how?

Accepted Answer

Progressively add layers from low to high resolution during training

Answer

Train all resolutions simultaneously with a single loss

Answer

Use a transformer instead of convolutions

Answer

Remove the discriminator at high resolutions

Question 12

What is the role of the mapping network and AdaIN in StyleGAN?

Accepted Answer

Mapping network produces a disentangled W; AdaIN injects per-layer styles

Answer

Mapping network classifies images; AdaIN is a loss term

Answer

They replace the discriminator with attention

Answer

They enforce cycle consistency

Question 13

Why was AdaIN replaced by weight demodulation in StyleGAN2, and what artifact did it fix?

Accepted Answer

Weight demodulation removes AdaIN-induced droplet/blob artifacts

Answer

It was replaced to add cycle consistency

Answer

It removed the need for a generator

Answer

It fixed exploding discriminator gradients

Question 14

What is the Fréchet Inception Distance (FID) and why is it preferred over Inception Score?

Accepted Answer

Fréchet distance between Gaussians fit to Inception features of real and fake images

Answer

The KL divergence between class predictions

Answer

The pixel-wise MSE between real and fake images

Answer

The discriminator accuracy on a held-out set

Question 15

What does the Inception Score (IS) measure and what is its main limitation?

Accepted Answer

It ignores the real data distribution and can miss intra-class mode collapse

Answer

It requires paired real/fake data

Answer

It directly measures pixel fidelity

Answer

It only works for conditional GANs

Question 16

What is spectral normalization and why is it effective for stabilizing GAN discriminators?

Accepted Answer

Divides weights by their largest singular value to bound the Lipschitz constant

Answer

Normalizes activations to zero mean and unit variance

Answer

Clips gradients to a fixed norm

Answer

Adds dropout to each layer

Question 17

Why are GANs notoriously difficult to train from a game-theory / optimization perspective?

Accepted Answer

It is a non-convex minimax game whose Nash equilibrium gradient methods may not reach

Answer

The loss is convex but has many local minima

Answer

There is no gradient signal at all

Answer

The networks share all weights

Question 18

What is feature matching as a GAN stabilization technique?

Accepted Answer

Match intermediate discriminator feature statistics of real and fake data

Answer

Match generator and discriminator weights

Answer

Match pixel histograms directly

Answer

Match the learning rates of both networks

Question 19

What is minibatch discrimination and what failure mode does it target?

Accepted Answer

Lets D compare samples across a batch to penalize low diversity (mode collapse)

Answer

Increases batch size for faster training

Answer

Normalizes features within a minibatch

Answer

Discriminates between minibatches and full batches

Question 20

How does the two-time-scale update rule (TTUR) help GAN convergence?

Accepted Answer

Use separate learning rates for G and D to reach a local Nash equilibrium

Answer

Use the same learning rate but two optimizers

Answer

Update G twice per D update always

Answer

Decay both learning rates to zero immediately

Question 21

What is the purpose of one-sided label smoothing in GAN training?

Accepted Answer

Soften real labels to ~0.9 to curb discriminator overconfidence

Answer

Set all labels to 0.5

Answer

Smooth only the fake labels to 0.1

Answer

Randomly flip all labels

Question 22

In Pix2Pix, why is the PatchGAN discriminator used instead of a full-image discriminator?

Accepted Answer

It scores local patches to capture high-frequency texture efficiently

Answer

It scores the whole image with a single scalar

Answer

It removes the need for an L1 loss

Answer

It enforces cycle consistency

Question 23

What is the difference between WGAN's critic and a standard GAN discriminator?

Accepted Answer

The critic outputs an unbounded Lipschitz-constrained score, not a 0-1 probability

Answer

The critic uses a softmax over classes

Answer

They are identical except for the name

Answer

The critic does not receive real samples

Question 24

What is the role of the latent space interpolation experiment in evaluating GANs?

Accepted Answer

Smooth interpolations indicate a learned manifold rather than memorization

Answer

It measures the FID directly

Answer

It is used to clip discriminator weights

Answer

It trains the mapping network

Question 25

How does an Auxiliary Classifier GAN (AC-GAN) differ from a standard conditional GAN?

Accepted Answer

The discriminator also predicts the class label via an auxiliary head

Answer

The generator predicts the discriminator's output

Answer

It removes class conditioning entirely

Answer

It uses two generators

Question 26

What is the vanishing gradient problem in the original GAN and how does it relate to JS divergence?

Accepted Answer

Disjoint supports make JS saturate, giving the generator zero gradient

Answer

The generator gradient explodes when distributions overlap

Answer

It is caused by too small a batch size only

Answer

It only occurs with the Wasserstein loss

Question 27

What is precision and recall in the context of generative model evaluation?

Accepted Answer

Precision = sample quality/realism; recall = coverage/diversity

Answer

Precision = diversity; recall = realism

Answer

Both measure pixel accuracy

Answer

They are only defined for classifiers, not GANs

Question 28

What is the purpose of the path length regularization introduced in StyleGAN2?

Accepted Answer

Encourage equal latent steps to cause equal-magnitude image changes (smoother mapping)

Answer

Penalize the discriminator gradient norm

Answer

Force the latent code to be sparse

Answer

Match generator and discriminator path lengths

Question 29

How does BigGAN scale GANs to high-quality ImageNet generation, and what is the truncation trick?

Accepted Answer

Truncating the latent at inference trades diversity for higher fidelity

Answer

Truncation removes the discriminator at test time

Answer

Truncation clips the generator weights

Answer

Truncation increases both diversity and fidelity for free

Question 30

What is GAN inversion and why is it important for image editing?

Accepted Answer

Finding the latent code that reconstructs a real image to enable editing

Answer

Inverting the discriminator to generate images

Answer

Reversing the training procedure

Answer

Computing the inverse of the weight matrices

Test yourself