Dilated and Atrous Convolutions Quiz

Answers and explanations appear as you take the quiz.

1. What does a dilation rate greater than 1 do to a convolution kernel?
   • It inserts gaps between the kernel weights so they cover a wider area
   • It adds more learnable weights to the kernel
   • It shrinks the kernel to a single point
   • It converts the kernel into a pooling operation
2. A 3x3 kernel with a dilation rate of 2 effectively spans how large a region?
   • 3x3
   • 5x5
   • 7x7
   • 9x9
3. Why are dilated convolutions especially useful for semantic segmentation?
   • They reduce the image to a single label
   • They eliminate the need for any training data
   • They enlarge the receptive field while keeping the feature map resolution high
   • They only work on grayscale images
4. What does the term atrous originate from?
   • A Latin word for layer
   • An acronym for a training method
   • The name of its inventor
   • The French a trous, meaning with holes
5. In WaveNet, why are stacked dilated convolutions valuable?
   • They generate images instead of audio
   • They capture very long-range dependencies in the audio waveform efficiently
   • They remove all silence from recordings
   • They convert audio into text directly
6. What problem can arise if you stack dilated convolutions with a poorly chosen single rate?
   • The network becomes too small to train
   • The image is permanently rotated
   • Gridding artifacts, where some input positions are never sampled
   • All weights become zero
7. How does Atrous Spatial Pyramid Pooling combine multiple scales in DeepLab?
   • It runs several dilation rates in parallel and fuses their outputs
   • It chooses one fixed dilation rate and discards the rest
   • It converts the image into a single global class before convolution
   • It replaces every convolution with a larger dense kernel
8. For one input channel and one output channel, how many kernel weights does a dilated 3×3 filter contain, excluding a bias?
   • 25, because dilation 2 spans a 5×5 region
   • 5, one for each position along the expanded width
   • 18, because dilation 2 doubles the number of weights
   • 9, because dilation changes spacing rather than the 3×3 weight count