Generative Model

Normalizing flows can only convert any (smooth) probability distribution into a Gaussian distribution.

• True

• False

Normalizing flows allow the exact evaluation of the probability (density) of a particular data sample.

• True

• False

Variational autoencoders (VAE) allow the exact evaluation of the probability (density) of a particular data sample.

• True

• False

Generative adversarial networks (GAN) allow the exact evaluation of the probability (density) of a particular data sample.

• True

• False

Normalizing flow models are trained by directly maximizing the log data likelihood.

• True

• False

Variational autoencoders (VAE) are trained by directly maximizing the log data likelihood.

• True

• False

Generative adversarial networks (GAN) are trained by directly maximizing the log data likelihood.

• True

• False

With normalizing flow models, the mapping between the visible and hidden variables is reversible.

• True

• False

Coupling layers have the properties that the log-determinant of the Jacobian matrix is easy to evaluate for back-propagation.

• True

• False

There is no constraint on how the input data should be split before being feed into the coupling layers.

• True

• False