#gan-properties

Question 1: Select all that apply

Which statements about the original GAN framework are true? Select all that apply.

The generator never writes down $p_g(x)$ ; it only produces samples1
The discriminator must be kept well synchronized with the generator during training2
No Markov chains are needed, neither to train the model nor to draw samples3
Drawing a sample from a trained generator requires running a Markov chain to convergence4
Training also yields an encoder that maps each data point back to its latent code $z$ 5