Depthwise Separable Convolutions Quiz

Answers and explanations appear as you take the quiz.

1. What are the two steps that a depthwise separable convolution splits a standard convolution into?
   • A pooling step followed by a normalization step
   • Two stacked 3x3 convolutions with a ReLU between them
   • A depthwise spatial filter per channel, then a 1x1 pointwise channel mix
   • An encoder convolution and a decoder deconvolution
2. In the depthwise step, how are the input channels treated?
   • Each channel is filtered independently with its own kernel
   • All channels are summed together first
   • Channels are randomly shuffled before filtering
   • Only the first channel is convolved
3. For many output channels, what operation-count reduction does the guide associate with a 3×3 depthwise separable convolution?
   • About 2 times less
   • About 100 times less
   • It uses more computation, not less
   • About 8 to 9 times less
4. What is the role of the pointwise (1x1) convolution in this design?
   • To downsample the spatial resolution
   • To combine information across channels at each location
   • To add positional encodings
   • To remove the bias terms
5. Which well-known architecture popularized depthwise separable convolutions for mobile devices?
   • MobileNet
   • ResNet
   • AlexNet
   • LSTM
6. An implementation applies only the depthwise spatial filters and omits the pointwise layer. Which capability is missing from that block?
   • Spatial filtering within each channel
   • Separate kernels for the input channels
   • Learned combination of information across channels
   • Local spatial feature extraction
7. Using the guide’s formulas with M = 4 input channels and N = 16 outputs, how many multiply-adds per location do standard and separable 3×3 convolutions require?
   • Standard 100; separable 576
   • Standard 576; separable 100
   • Standard 576; separable 36
   • Standard 64; separable 100
8. In the guide’s cost ratio 1/N + 1/9 for a 3×3 separable convolution, what happens as the output-channel count N becomes very large?
   • The ratio approaches 9
   • The ratio approaches zero
   • The ratio becomes independent of kernel size for every kernel
   • The ratio approaches 1/9