U-Net Architecture Quiz

Answers and explanations appear as you take the quiz.

1. What gives U-Net its 'U' shape?
   • An encoder that downsamples and a symmetric decoder that upsamples
   • A single straight stack of convolutions
   • Two parallel networks with no connection
   • A recurrent loop over time steps
2. What is the purpose of U-Net's skip connections?
   • To reduce the number of parameters to zero
   • To pass fine spatial detail from encoder layers directly to matching decoder layers
   • To convert the image to grayscale
   • To skip training entirely
3. What was U-Net originally designed for?
   • Speech recognition
   • Stock price prediction
   • Biomedical image segmentation
   • Text translation
4. In Stable Diffusion, what does the U-Net predict at each step?
   • The final caption of the image
   • The compression ratio
   • The camera angle
   • The noise to remove from the noisy latent
5. What happens to spatial information as the encoder downsamples?
   • It abstracts 'what' is present but loses precise 'where' detail
   • It gains exact pixel locations
   • It converts the image to text
   • It doubles the resolution
6. Which augmentation did the original U-Net paper use to learn invariance from limited annotated biomedical images?
   • Replacing every image with a random label
   • Elastic deformations of the training images
   • Training only on unmodified test images
   • Removing the decoder path before training
7. How did the original U-Net paper process large images beyond the network’s input size?
   • By discarding all image borders
   • By resizing every image to one pixel
   • By predicting on overlapping tiles and extrapolating context at tile borders
   • By running the model once on a text prompt
8. What did the original U-Net paper’s weighted loss emphasize?
   • Only background pixels, regardless of distance
   • The most common image color
   • Noise prediction at a diffusion timestep
   • Border pixels between touching cells to help separate adjacent objects