#rbm

Question 1: Single choice

Restricted Boltzmann machines and variational autoencoders are both latent-variable models whose likelihood $P(X)$ is intractable. How does each cope with it?

RBMs use Gibbs sampling; VAEs fit an encoder $Q_\theta(z \mid X)$ to the posterior1
Both use Gibbs sampling to approximate the gradient of $\log P(X)$ 2
Both fit an encoder $Q_\theta(z \mid X)$ to the posterior and maximise a lower bound3
Neither needs to approximate, since both compute $P(X)$ in closed form4

Timed test