Spatial Transformer Networks Quiz

Answers and explanations appear as you take the quiz.

1. What capability does a Spatial Transformer Network add to a neural network?
   • The ability to learn geometric transformations of its input, like rotation and scaling
   • The ability to compress images losslessly
   • The ability to generate text captions
   • The ability to remove all noise from audio
2. Which three components make up a spatial transformer module?
   • Encoder, decoder, and discriminator
   • Localization network, grid generator, and sampler
   • Convolution, pooling, and softmax
   • Embedding, attention, and feed-forward
3. Why can a Spatial Transformer Network be trained with ordinary backpropagation?
   • Because it uses no parameters
   • Because it only runs at inference time
   • Because every step, including bilinear sampling, is differentiable
   • Because it is trained with reinforcement learning
4. What does the localization network typically output for an affine transformation?
   • A single confidence score
   • A list of bounding boxes
   • A segmentation mask
   • A 2x3 transformation matrix of parameters
5. Why are standard CNNs only weakly invariant to spatial changes, motivating STNs?
   • They rely mainly on pooling for limited spatial tolerance
   • They never use convolutions
   • They cannot process images larger than 28x28
   • They require labeled transformations for every image
6. Which application is a natural fit for a Spatial Transformer Network?
   • Sorting numbers in a database
   • Straightening tilted or curved text before recognition
   • Compressing a video file
   • Generating random noise
7. What does the grid generator do in a spatial transformer?
   • Assigns a class label to every image pixel
   • Measures the confidence of each object detection
   • Maps output pixel locations back to source-image coordinates using predicted transform parameters
   • Changes the neural network’s vocabulary
8. What interpolation method does the spatial transformer sampler use?
   • Beam search
   • Nearest-neighbor word matching
   • Fourier phase rotation only
   • Bilinear interpolation