Swin Transformer Quiz

Answers and explanations appear as you take the quiz.

1. What does the 'Swin' in Swin Transformer stand for?
   • Shifted Windows
   • Smooth Window
   • Spatial Winding
   • Single Window
2. What positional information does Swin add within each local attention window?
   • Absolute word embeddings from a text tokenizer
   • A learnable relative-position bias
   • A global convolution kernel shared across all windows
   • A class-specific bounding-box regression target
3. How does Swin let information move between separate windows?
   • It uses recurrent connections
   • It deletes window boundaries entirely
   • It shifts the window partition in alternating layers
   • It only uses a single global window
4. What do patch-merging layers do between Swin stages?
   • Increase resolution and remove channels
   • Randomly drop patches
   • Translate patches into words
   • Merge neighboring patches, reducing resolution and increasing channels
5. Why is Swin's hierarchical, multi-scale output especially useful?
   • It plugs neatly into detection and segmentation frameworks that expect feature pyramids
   • It only helps with single-label classification
   • It removes the need for labeled data
   • It makes the model smaller than any CNN
6. How does Swin’s attention pattern differ from standard ViT attention?
   • Swin has no attention mechanism
   • Swin computes attention in local windows and shifts partitions between blocks; standard ViT uses global patch attention
   • ViT uses only local convolutions
   • Both always use identical global attention
7. What scaling property does fixed-size window attention provide in Swin?
   • It is independent of image resolution
   • It becomes quadratic in channel count only
   • Attention computation grows linearly with image size for fixed window size
   • It removes all attention computation
8. Why does shifted-window attention use a mask after the cyclic shift?
   • To force every patch to attend to every image patch
   • To remove all positional information
   • To convert window attention into a convolution
   • To prevent tokens from attending across boundaries that should remain separate windows