Convolutional Neural Networks Quiz

Answers and explanations appear as you take the quiz.

1. What is the main job of a filter (kernel) in a CNN?
   • To store the final classification label
   • To randomly shuffle the pixels before training
   • To slide across the image and detect a specific visual pattern
   • To convert the image into audio
2. How does spatial kernel sharing reduce parameter requirements compared with learning a separate kernel at every image location?
   • The same learned kernel weights are reused across locations
   • All input pixels are deleted
   • Every layer must use only grayscale data
   • The model avoids all training
3. What does a spatial pooling layer commonly do to a CNN feature map?
   • Assign a final label to every object
   • Reduce its spatial resolution
   • Guarantee invariance to every possible image shift
   • Always increase the number of input channels
4. In a deep CNN, how do the features generally change from early layers to later layers?
   • From whole objects to single pixels
   • They stay identical at every layer
   • From text to numbers only
   • From simple edges and colors to complex parts like eyes or wheels
5. Which named CNN is associated with the influential 2012 ImageNet classification result?
   • AlexNet
   • ResNet
   • Vision Transformer
   • SqueezeNet
6. What numeric operation happens when a filter is applied to a patch of the image?
   • The pixels are sorted from darkest to brightest
   • The patch is deleted from the image
   • Each weight is multiplied by its overlapping pixel, and the products are summed
   • The pixels are converted to letters
7. With a 6-by-6 input and 3-by-3 kernel at stride one, no padding, and no dilation, what is the CNN output-grid size?
   • 6 by 6
   • 3 by 3
   • 2 by 2
   • 4 by 4
8. In dense ungrouped convolution on a three-channel input, which channels does one filter span?
   • Only the first channel by definition
   • All three input channels
   • No input channels because kernels contain only output labels
   • Exactly one channel in every convolution configuration