CS7643 Quiz 4 – Deep Learning Exam (50
Questions with Answers & Rationales) Georgia
Tech
1. What is the main purpose of convolution in a CNN?
Answer: To detect local spatial features such as edges or textures.
Rationale: Convolutions slide filters across input data to extract hierarchical spatial features.
2. Which operation reduces the spatial size of feature maps?
Answer: Pooling.
Rationale: Pooling layers (e.g., max pooling) downsample feature maps to reduce computation and
overfitting.
3. Why is ReLU used as an activation function in CNNs?
Answer: It introduces nonlinearity and mitigates the vanishing gradient problem.
Rationale: ReLU keeps positive gradients constant, accelerating training.
4. What does “stride” mean in convolution?
Answer: The number of pixels the filter moves per step.
Rationale: Larger strides reduce output dimensions and computation.
5. Batch normalization primarily helps to:
Answer: Stabilize and speed up training.
Rationale: It normalizes activations within each mini-batch, improving gradient flow.
6. Why does dropout reduce overfitting?
Answer: It randomly deactivates neurons during training.
Rationale: This prevents units from co-adapting and improves model generalization.
, 7. The output size of a convolutional layer depends on:
Answer: Input size, kernel size, stride, and padding.
Rationale: These four parameters determine output feature map dimensions.
8. What is transfer learning?
Answer: Using a pre-trained model on a new but related task.
Rationale: It leverages learned features to reduce data and computation requirements.
9. What is the main benefit of fine-tuning a pre-trained network?
Answer: It adapts existing feature representations to a specific new dataset.
Rationale: Fine-tuning retains low-level features while refining higher layers.
10. L2 regularization discourages:
Answer: Large weights.
Rationale: It penalizes weight magnitude, promoting smaller, smoother parameter values.
11. Early stopping is a form of:
Answer: Regularization.
Rationale: It halts training before overfitting occurs on validation data.
12. What does the Adam optimizer combine?
Answer: Momentum and adaptive learning rates.
Rationale: Adam integrates RMSProp and Momentum for faster, stable convergence.
13. What does a receptive field represent in CNNs?
Answer: The input area influencing a neuron’s activation.
Rationale: Larger receptive fields capture more global context.
14. Gradient clipping is used to prevent:
Answer: Exploding gradients.
Rationale: It limits gradient magnitude during backpropagation.
15. What happens when learning rate is too high?
Answer: The model may diverge or oscillate.
Rationale: Updates overshoot optimal parameters.
Questions with Answers & Rationales) Georgia
Tech
1. What is the main purpose of convolution in a CNN?
Answer: To detect local spatial features such as edges or textures.
Rationale: Convolutions slide filters across input data to extract hierarchical spatial features.
2. Which operation reduces the spatial size of feature maps?
Answer: Pooling.
Rationale: Pooling layers (e.g., max pooling) downsample feature maps to reduce computation and
overfitting.
3. Why is ReLU used as an activation function in CNNs?
Answer: It introduces nonlinearity and mitigates the vanishing gradient problem.
Rationale: ReLU keeps positive gradients constant, accelerating training.
4. What does “stride” mean in convolution?
Answer: The number of pixels the filter moves per step.
Rationale: Larger strides reduce output dimensions and computation.
5. Batch normalization primarily helps to:
Answer: Stabilize and speed up training.
Rationale: It normalizes activations within each mini-batch, improving gradient flow.
6. Why does dropout reduce overfitting?
Answer: It randomly deactivates neurons during training.
Rationale: This prevents units from co-adapting and improves model generalization.
, 7. The output size of a convolutional layer depends on:
Answer: Input size, kernel size, stride, and padding.
Rationale: These four parameters determine output feature map dimensions.
8. What is transfer learning?
Answer: Using a pre-trained model on a new but related task.
Rationale: It leverages learned features to reduce data and computation requirements.
9. What is the main benefit of fine-tuning a pre-trained network?
Answer: It adapts existing feature representations to a specific new dataset.
Rationale: Fine-tuning retains low-level features while refining higher layers.
10. L2 regularization discourages:
Answer: Large weights.
Rationale: It penalizes weight magnitude, promoting smaller, smoother parameter values.
11. Early stopping is a form of:
Answer: Regularization.
Rationale: It halts training before overfitting occurs on validation data.
12. What does the Adam optimizer combine?
Answer: Momentum and adaptive learning rates.
Rationale: Adam integrates RMSProp and Momentum for faster, stable convergence.
13. What does a receptive field represent in CNNs?
Answer: The input area influencing a neuron’s activation.
Rationale: Larger receptive fields capture more global context.
14. Gradient clipping is used to prevent:
Answer: Exploding gradients.
Rationale: It limits gradient magnitude during backpropagation.
15. What happens when learning rate is too high?
Answer: The model may diverge or oscillate.
Rationale: Updates overshoot optimal parameters.
