CS 7643 EXAM 2 ACTUAL 2026/2027 HIGH YIELD
PRACTICE QUESTIONS AND STUDY GUIDE
ACCURATE EXAM
Section 1: CNNs
Q1. What is the primary purpose of a convolutional layer?
A. Flatten spatial dimensions
B. Extract local features while preserving spatial relationships
C. Replace activation functions
D. Increase batch size
Answer: B
Rationale: Convolution learns local patterns and preserves spatial structure.
Q2. Input 32×32, kernel 5×5, stride 1, padding 0. Output size?
A. 28×28
B. 32×32
C. 27×27
D. 30×30
Answer: A
Rationale: (32−5+0)/1+1=28(32−5+0)/1+1=28.
Q3. “Same” padding with stride 1 does what?
A. Shrinks output
B. Preserves input spatial size
C. Doubles channels
D. Removes bias
Answer: B
Rationale: Same padding keeps output width/height equal to input.
Q4. Conv layer: 3 input channels, 64 filters, 3×3 kernel, bias. Parameter count?
A. 1,728
B. 1,792
C. 576
D. 64
,Answer: B
Rationale: 64×(3×3×3+1)=179264×(3×3×3+1)=1792.
Q5. A 1×1 convolution is mainly used to:
A. Change spatial size
B. Change channel depth and add nonlinearity
C. Pool features
D. Flip the kernel
Answer: B
Rationale: 1×1 conv mixes channels without changing spatial dimensions.
Q6. Max pooling primarily provides:
A. Translation invariance
B. More learnable parameters
C. Kernel flipping
D. Channel expansion
Answer: A
Rationale: Max pooling makes features more robust to small shifts.
Q7. Receptive field means:
A. Number of filters
B. Input region affecting a neuron
C. Batch size
D. Learning rate
Answer: B
Rationale: It is the region of input that influences a particular activation.
Q8. Two stacked 3×3 convs vs one 5×5 conv: main advantage?
A. Fewer parameters and more nonlinearity
B. More parameters
C. No nonlinearity
D. Larger stride
Answer: A
Rationale: Two 3×3 convs use 18 weights vs 25 and add two ReLUs.
Q9. Weight sharing in CNNs means:
A. Same filter applied across spatial locations
B. Unique weight per pixel
, C. No bias
D. Random weights
Answer: A
Rationale: Weight sharing reduces parameters and gives translation equivariance.
Q10. Stride greater than 1 causes:
A. Downsampling
B. Upsampling
C. Same output size
D. More channels
Answer: A
Rationale: Larger stride skips positions, reducing spatial dimensions.
Q11. “Valid” padding means:
A. No padding
B. Same padding
C. Half padding
D. Zero padding always
Answer: A
Rationale: Valid padding uses no extra border zeros.
Q12. Global average pooling does what?
A. Averages each feature map to one value
B. Max pools each feature map
C. Flattens all channels
D. Upsamples feature maps
Answer: A
Rationale: It reduces each channel to a single scalar.
Q13. Dilated convolution increases:
A. Receptive field without increasing parameters
B. Number of channels
C. Batch size
D. Stride only
Answer: A
Rationale: Dilation spaces kernel elements to enlarge receptive field.
PRACTICE QUESTIONS AND STUDY GUIDE
ACCURATE EXAM
Section 1: CNNs
Q1. What is the primary purpose of a convolutional layer?
A. Flatten spatial dimensions
B. Extract local features while preserving spatial relationships
C. Replace activation functions
D. Increase batch size
Answer: B
Rationale: Convolution learns local patterns and preserves spatial structure.
Q2. Input 32×32, kernel 5×5, stride 1, padding 0. Output size?
A. 28×28
B. 32×32
C. 27×27
D. 30×30
Answer: A
Rationale: (32−5+0)/1+1=28(32−5+0)/1+1=28.
Q3. “Same” padding with stride 1 does what?
A. Shrinks output
B. Preserves input spatial size
C. Doubles channels
D. Removes bias
Answer: B
Rationale: Same padding keeps output width/height equal to input.
Q4. Conv layer: 3 input channels, 64 filters, 3×3 kernel, bias. Parameter count?
A. 1,728
B. 1,792
C. 576
D. 64
,Answer: B
Rationale: 64×(3×3×3+1)=179264×(3×3×3+1)=1792.
Q5. A 1×1 convolution is mainly used to:
A. Change spatial size
B. Change channel depth and add nonlinearity
C. Pool features
D. Flip the kernel
Answer: B
Rationale: 1×1 conv mixes channels without changing spatial dimensions.
Q6. Max pooling primarily provides:
A. Translation invariance
B. More learnable parameters
C. Kernel flipping
D. Channel expansion
Answer: A
Rationale: Max pooling makes features more robust to small shifts.
Q7. Receptive field means:
A. Number of filters
B. Input region affecting a neuron
C. Batch size
D. Learning rate
Answer: B
Rationale: It is the region of input that influences a particular activation.
Q8. Two stacked 3×3 convs vs one 5×5 conv: main advantage?
A. Fewer parameters and more nonlinearity
B. More parameters
C. No nonlinearity
D. Larger stride
Answer: A
Rationale: Two 3×3 convs use 18 weights vs 25 and add two ReLUs.
Q9. Weight sharing in CNNs means:
A. Same filter applied across spatial locations
B. Unique weight per pixel
, C. No bias
D. Random weights
Answer: A
Rationale: Weight sharing reduces parameters and gives translation equivariance.
Q10. Stride greater than 1 causes:
A. Downsampling
B. Upsampling
C. Same output size
D. More channels
Answer: A
Rationale: Larger stride skips positions, reducing spatial dimensions.
Q11. “Valid” padding means:
A. No padding
B. Same padding
C. Half padding
D. Zero padding always
Answer: A
Rationale: Valid padding uses no extra border zeros.
Q12. Global average pooling does what?
A. Averages each feature map to one value
B. Max pools each feature map
C. Flattens all channels
D. Upsamples feature maps
Answer: A
Rationale: It reduces each channel to a single scalar.
Q13. Dilated convolution increases:
A. Receptive field without increasing parameters
B. Number of channels
C. Batch size
D. Stride only
Answer: A
Rationale: Dilation spaces kernel elements to enlarge receptive field.
