A data scientist is training a deep CNN for image classification. The training loss decreases rapidly but then plateaus, and the validation loss starts to increase. The model uses ReLU activations throughout. Which change is most likely to improve generalization?

Replace ReLU with sigmoid activations in all layers.

Remove batch normalization layers to allow the model to fit the training data better.

Add dropout layers after the convolutional and fully connected layers.

Increase the number of filters in each convolutional layer.

During training of a CNN, the loss function decreases but the gradients in the early layers become very small, and the network stops learning. The network uses sigmoid activations in the hidden layers. What is the most likely cause and the best fix?

The learning rate is too high; reduce the learning rate.

Sigmoid activations cause vanishing gradients; replace them with ReLU activations.

The network is overfitting; add dropout.

The batch size is too small; increase the batch size.

What is the effect of increasing the number of filters in a convolutional layer?

It reduces the risk of overfitting by simplifying the model.

It speeds up training because fewer iterations are needed.

It increases the number of feature maps and the number of learnable parameters.

It decreases the spatial size of the output feature maps.

A startup has a small dataset of 5,000 labeled images of industrial defects. They need to build a classifier that achieves high accuracy quickly. Which approach is most appropriate?

Use a pretrained model but freeze all layers and use it as a fixed feature extractor without any fine-tuning.

Train a deep CNN from scratch with random initialization.

Use a pretrained ResNet model and fine-tune only the final fully connected layers.

Use a simple linear classifier on raw pixel values.

GIAC Machine Learning Engineer

Domain 3Objective 2

Convolutional Neural Networks GMLE Practice Questions (Page 6)

Part of the Advanced Machine Learning and Neural Networks domain, which makes up ~30% of our current practice bank. GIAC (SANS) does not publish an official question count, but from its 180-minute exam (~70–120 total, ~21–36 in this domain), expect 7–12 from this objective — we provide 40 practice questions to prepare you well beyond it. (estimate)

40questions here

8free pages

10concepts

Questions 26–30

26application · medium
A data scientist is training a deep CNN for image classification. The training loss decreases rapidly but then plateaus, and the validation loss starts to increase. The model uses ReLU activations throughout. Which change is most likely to improve generalization?
Select an answer first
27foundation · easy
When applying a 3x3 convolution filter with stride 1 and no padding to a 5x5 input, what is the spatial size of the resulting feature map?
Select an answer first
28application · medium
During training of a CNN, the loss function decreases but the gradients in the early layers become very small, and the network stops learning. The network uses sigmoid activations in the hidden layers. What is the most likely cause and the best fix?
Select an answer first
29foundation · easy
What is the effect of increasing the number of filters in a convolutional layer?
Select an answer first
30application · medium
A startup has a small dataset of 5,000 labeled images of industrial defects. They need to build a classifier that achieves high accuracy quickly. Which approach is most appropriate?
Select an answer first

Finished these 5 questions?

Review the revealed explanations, or continue through the curriculum.

Next 5 questions →Back to GMLE

Free Basic Practice is a study aid with revealable answers — not a scored exam. Examers.io is independent and not affiliated with or endorsed by GIAC (SANS). “GMLE” is a trademark of its owner, used for identification only.