Theory test: Generating images

6 questions. Pass with 5 correct.

1. 01 What is the reparameterisation trick for?

   Speeding up the decoderMaking sampling from $\mathcal{N}(\mu ,{\sigma }^2)$ differentiable with respect to $\mu$ and $\sigma$Removing the KL termChoosing the latent size
2. 02 What does the KL term in a VAE's loss do?

   Improves reconstruction sharpnessPulls each input's latent distribution toward $\mathcal{N}(0,I)$, making the latent space smooth and samplableMeasures classification accuracyPrevents the decoder from overfitting to noise
3. 03 A GAN's discriminator is trained to:

   Generate imagesClassify inputs as real or generatedPredict the noise in an imageEncode images to latents
4. 04 Mode collapse in a GAN means:

   The discriminator stops learningThe generator produces only a narrow variety of outputsThe loss becomes NaNTraining takes too long
5. 05 In DDPM, the denoising network is trained to predict:

   The class labelThe noise $ϵ$ that was added to the imageThe next pixelThe step size ${\beta }_t$
6. 06 When ${\bar{\alpha }}_T\approx 0$, the final noised image $x_T$ is:

   The original imageAlmost pure Gaussian noiseA blurred version of the imageAll zeros