Theory test: Convolutional networks

6 questions. Pass with 5 correct.

1. 01 A 32×32 image goes through a 5×5 convolution, stride 1, no padding. The output is:

   32×3228×2827×2730×30
2. 02 A convolutional layer maps 3 input channels to 16 output channels with 3×3 kernels. How many parameters (with biases)?

   14443244827
3. 03 What does parameter sharing mean in a CNN?

   Different layers share weightsThe same kernel weights are applied at every spatial positionTwo networks are trained togetherChannels share one bias
4. 04 Two stacked 3×3 convolutions (stride 1) give each output unit a receptive field of:

   3×35×56×69×9
5. 05 Why do residual connections help very deep networks train?

   They reduce the number of parametersThe identity path lets gradients flow directly and lets blocks default to doing nothingThey replace the need for activationsThey add dropout
6. 06 2×2 max pooling with stride 2 on a 28×28 feature map gives:

   27×2714×147×726×26