Convolutional Neural Networks Quiz

An input volume of size 64x64x3 is passed through a convolutional layer with 16 filters, each of size 3x3, a stride of 1, and 'SAME' padding. What are the dimensions of the output volume?

• 62x62x16

• 64x64x16

• 64x64x3

• 32x32x16

An input volume of size 32x32x3 is passed through a convolutional layer with 10 filters, each of size 5x5, a stride of 1, and 'VALID' (zero) padding. What are the dimensions of the output volume?

• 32x32x10

• 28x28x10

• 27x27x10

• 32x32x3

An input volume of size 20x20x64 is passed through a Max Pooling layer with a 2x2 filter and a stride of 2. What are the dimensions of the output volume?

• 20x20x32

• 10x10x32

• 10x10x64

• 19x19x64

A convolutional layer has an input volume of 64x64x16. It uses 32 filters, each of size 3x3. What is the total number of learnable parameters (weights and biases) in this layer?

• 4,608

• 4,640

• 144

• 1,536

Consider a convolutional layer with 64 filters of size 3x3x32 (HxWxChannels_in). What is the total number of weight parameters in this layer?

• 576

• 18,432

• 18,496

• 2,048

A feature map of size 14x14x128 is passed to a fully connected (Dense) layer with 1024 neurons. How many weight parameters are there in this connection, assuming the feature map is flattened first?

• (14 * 128) + 1024

• (14 * 14 * 128) * 1024

• (14 + 14 + 128) * 1024

• 14 * 14 * 128 * 1024 + 128

An input volume of 7x7x512 is the input to two successive convolutional layers. Layer 1 has 128 filters of 1x1, stride 1. Layer 2 has 256 filters of 3x3, stride 1, 'SAME' padding. How many parameters are in Layer 2?

• (11512 + 1) * 128

• (33512 + 1) * 256

• (33128 + 1) * 256

• (11128 + 1) * 256

You want to design a convolutional layer that downsamples a 56x56 feature map to 28x28. Which of the following configurations would achieve this?

• 3x3 filter, stride 1, 'SAME' padding.

• 3x3 filter, stride 2, 'SAME' padding.

• 2x2 filter, stride 1, 'VALID' padding.

• 5x5 filter, stride 1, 'VALID' padding.

A convolutional layer is defined with a 7x7 filter, a stride of 2, and 3 units of padding. If the input width is 32, what is the output width?

• 15

• 16

• 17

• 18

A "bottleneck" residual block in ResNet-50 takes a 256-channel feature map, applies a 1x1 conv with 64 filters, then a 3x3 conv with 64 filters, and finally a 1x1 conv with 256 filters. Why is this more efficient than a single 3x3 conv with 256 filters?

• It has a larger receptive field.

• The total number of parameters is significantly lower.

• It uses fewer layers.

• It allows for a larger stride.