CS7643 QUIZ 1 QUESTIONS WITH DETAILED
VERIFIED ANSWERS (100% CORRECT
ANSWERS) /ALREADY GRADED
Convolution Features
edges
colors
textures
motifs (corners, shapes)
Receptive field
A region of an image (image patch) from which the node receives input. Usually denoted by a
K1 x K2 matrix.
Convolution vs Cross-correlation
Convolution: flip the kernel (rotate 180) and take the dot product with image patch
Cross-correlation: do not flip the kernel to take the dot product with image patch
Advantage of using image patch
1./ Reduces the input parameters to
K1 x K2 + 1 (bias)
for each output node. Thus, the total number of input parameters:
N x (K1 + K2 + 1)
2./ Explicitly maintains spatial information
Weight sharing
, The weights will represent what types of features we will extract. The weights (W) will be the
same for each output node with respect to a specific kernel, regardless of the specific image
patch we are looking at.
The total number of input parameters:
K1 x K2 + 1
Input parameters with multiple feature extractions
(K1 x K2 + 1) x M
where M is the number of features
Relationship between convolution and cross-correlation
Duality: If cross-correlation is the forward pass (which is the easier operation), the convolution
operation is going to be the backward pass to calculate gradients (vice versa)
Valid convolution
When the kernel is fully on the image. (No padding)
Output size of the vanilla convolution,
given H, W, K1, K2
(H - K1 + 1) x (W - K2 + 1)
How to add padding
Increases the size of the image with P in both directions (top & bottom, left & right)
--> (H + 2P) x (W + 2P)
Can be filled with zeros or mirror the image
VERIFIED ANSWERS (100% CORRECT
ANSWERS) /ALREADY GRADED
Convolution Features
edges
colors
textures
motifs (corners, shapes)
Receptive field
A region of an image (image patch) from which the node receives input. Usually denoted by a
K1 x K2 matrix.
Convolution vs Cross-correlation
Convolution: flip the kernel (rotate 180) and take the dot product with image patch
Cross-correlation: do not flip the kernel to take the dot product with image patch
Advantage of using image patch
1./ Reduces the input parameters to
K1 x K2 + 1 (bias)
for each output node. Thus, the total number of input parameters:
N x (K1 + K2 + 1)
2./ Explicitly maintains spatial information
Weight sharing
, The weights will represent what types of features we will extract. The weights (W) will be the
same for each output node with respect to a specific kernel, regardless of the specific image
patch we are looking at.
The total number of input parameters:
K1 x K2 + 1
Input parameters with multiple feature extractions
(K1 x K2 + 1) x M
where M is the number of features
Relationship between convolution and cross-correlation
Duality: If cross-correlation is the forward pass (which is the easier operation), the convolution
operation is going to be the backward pass to calculate gradients (vice versa)
Valid convolution
When the kernel is fully on the image. (No padding)
Output size of the vanilla convolution,
given H, W, K1, K2
(H - K1 + 1) x (W - K2 + 1)
How to add padding
Increases the size of the image with P in both directions (top & bottom, left & right)
--> (H + 2P) x (W + 2P)
Can be filled with zeros or mirror the image
