Variational Autoencoders Quiz

Answers and explanations appear as you take the quiz.

1. What does the encoder in a VAE output for a given input?
   • The parameters of a probability distribution, such as a mean and variance
   • A single fixed point in latent space
   • A class label
   • A reconstructed copy of the input
2. What is the purpose of the reparameterization trick?
   • To make the decoder larger
   • To allow gradients to flow through the random sampling step
   • To remove the need for a decoder
   • To compress the latent space further
3. In the standard Gaussian VAE formulation, what does the KL term encourage?
   • The decoder to predict future input tokens
   • The encoder to remove all randomness
   • The approximate posterior to stay close to the chosen prior, commonly a standard normal
   • Every latent variable to have the same observed value
4. The objective function that VAEs maximize during training is called the:
   • Cross-entropy ceiling
   • Adversarial loss
   • Mean squared margin
   • Evidence Lower Bound (ELBO)
5. In a Gaussian VAE, what enables generation of a new sample?
   • Sample a latent from the prior and decode it with the learned decoder
   • Copy a training image directly without a model
   • Run the encoder backward with no latent sampling
   • Choose a random class label without decoding
6. In a latent diffusion model like Stable Diffusion, what role does the VAE play?
   • It labels the training images
   • It compresses images to a latent space where diffusion runs, then decodes back to pixels
   • It replaces the text encoder
   • It schedules the noise levels
7. In variational inference, what does the ELBO provide for the marginal log likelihood log p(x)?
   • An exact value of log p(x) for every model
   • A guarantee that each generated sample is realistic
   • A tractable lower bound that can be maximized instead of directly maximizing log p(x)
   • A posterior distribution that requires no inference model
8. What distribution does a VAE decoder model given a latent vector?
   • The encoder’s approximate posterior q(z|x)
   • The training set’s class-frequency histogram only
   • The optimizer’s learning-rate schedule
   • The likelihood of an observation conditioned on that latent, p(x|z)