VQ-VAE and Discrete Latents Quiz

Answers and explanations appear as you take the quiz.

1. What makes VQ-VAE's latent space different from a standard VAE's?
   • It is discrete, using a learned codebook of embeddings
   • It is continuous and Gaussian
   • It has no latent space at all
   • It stores raw pixels
2. How does VQ-VAE pass gradients through the non-differentiable quantization step?
   • It skips backpropagation entirely
   • With a straight-through estimator that copies gradients past the quantizer
   • By using reinforcement learning only
   • By rounding gradients to integers
3. What is 'codebook collapse'?
   • The decoder ignores the latents
   • The encoder outputs become infinite
   • Only a small fraction of codebook entries ever get used
   • The image resolution drops to zero
4. Why are discrete latents useful for generative modeling with Transformers?
   • They make images smaller on disk only
   • They eliminate the need for a decoder
   • They remove all randomness from generation
   • They turn media into token sequences a Transformer can model like language
5. What did VQGAN add on top of the basic VQ-VAE recipe?
   • Adversarial and perceptual losses for sharper reconstructions
   • A diffusion sampler
   • Removal of the codebook
   • A recurrent encoder
6. Which loss term keeps the encoder 'committed' to the codebook entries it selects?
   • The KL divergence
   • The commitment loss
   • The cross-entropy loss
   • The triplet loss
7. What kind of prior does the original VQ-VAE paper learn over its discrete codes?
   • A fixed prior that is never learned
   • A prior over camera locations only
   • A learned prior over the discrete code sequence
   • No prior or decoder distribution
8. Which text-to-image paper used a discrete VAE to tokenize images before autoregressive generation?
   • The original Whisper speech-recognition paper
   • A paper on Kalman-filter vehicle tracking
   • The original BERT masked-language-model paper
   • The original DALL·E paper