#vae

Question 1: Single choice

You train a plain convolutional autoencoder and a VAE with the same $K$ -dimensional bottleneck on MNIST, then feed both decoders random codes $\mathbf{z} \sim \mathcal{N}(\mathbf{0}, I)$ . Only the VAE's outputs look like handwritten digits. Why?

Its KL term keeps codes near the prior $\mathcal{N}(0, I)$ , so prior samples fall where the decoder was trained1
The VAE decoder has more capacity, so it can turn any random code into a plausible-looking digit2
The plain decoder ignores its input after training and always outputs the average training image3
Only the VAE rescales its decoder outputs to the range of real pixel values, between 0 and 14

Timed test