Generative Adversarial Networks

Question 1 of 5

Under the taxonomy of generative models presented in the lecture, GANs are classified as which type of density model?

Explicit density model with tractable densityExplicit density model with approximate densityImplicit density modelAutoregressive density model

Question 2 of 5

Why is the original GAN generator loss 𝔼[log(1−D(G(z)))] problematic in practice, and what replacement is commonly used?

It leads to mode collapse because the generator ignores certain modes; the fix is to add a reconstruction loss.Its gradient saturates when the discriminator is confident, providing no learning signal; instead, the generator maximizes 𝔼[log D(G(z))].It causes the discriminator to overfit to real data; the solution is to use label smoothing.It makes training unstable because the loss is non-convex; the replacement is the least-squares loss from LSGAN.

Question 3 of 5

What does a linear walk between two latent vectors in the generator's input space typically demonstrate?

That the discriminator's loss value correlates strongly with the quality of generated samplesThat the generator has learned a smooth manifold rather than memorizing training pointsThat the generator can produce images at arbitrarily high resolutions without additional trainingThat the generator and discriminator have reached a global optimum of the minimax game

Question 4 of 5

Why does the KL divergence often fail to provide useful gradients when training high-dimensional image generators, and what distance does the Wasserstein GAN use instead?

KL divergence is zero when distributions are disjoint; WGAN uses the Jensen-Shannon divergence.KL divergence is infinite when the distributions' supports are disjoint; WGAN uses the Earth Mover distance.KL divergence is only defined for discrete distributions; WGAN uses a continuous relaxation.KL divergence requires an invertible generator; instead, WGAN uses a critic network to approximate the divergence.

Question 5 of 5

What is the key modification that InfoGAN makes to the standard GAN framework to learn disentangled representations?

It adds an auxiliary classifier to the discriminator that predicts the class labels.It maximizes mutual information between a subset of the latent variables and the generator's output.It uses a variational autoencoder as the generator to enforce a structured latent space.It conditions the generator on additional labels provided during training, making it a conditional GAN.