Convolutional Neural Networks GMLE Practice Questions (Page 5)

Questions 21–25

1. 21foundation · easy

   Which statement best describes the role of the fully connected layers in a CNN designed for image classification?

   • AThey combine the high-level features extracted by earlier layers to produce class scores.
   • BThey apply a fixed nonlinear transformation to each pixel independently.
   • CThey learn to detect low-level features such as edges and corners in the input image.
   • DThey reduce the spatial dimensions of the feature maps to control overfitting.
   Reveal Solution Discussion Select an answer first
2. 22application · medium

   A research team has a large dataset of 1 million images of animals and wants to build a model to classify 100 species. They have access to a pretrained VGG16 model. What is the most effective way to use this model?

   • AFine-tune the entire VGG16 model on the new dataset.
   • BReplace the final fully connected layer with a new layer and train only that layer.
   • CFreeze all layers and use the pretrained model as a fixed feature extractor.
   • DUse the pretrained model without any modification and directly classify the 100 species.
   Reveal Solution Discussion Select an answer first
4. 23expert · hard

   A team needs to build a model that can detect and classify multiple objects in an image, but they also need to provide a segmentation mask for each object. They have a dataset with both bounding boxes and pixel-level masks. Which architecture is most appropriate?

   • AA standard CNN classifier with a softmax output
   • BYOLO with an additional fully connected layer
   • CU-Net with a classification head
   • DMask R-CNN with a pretrained backbone
   Reveal Solution Discussion Select an answer first
5. 24foundation · easy

   When fine-tuning a pretrained CNN for a new task with a different number of classes, what must be modified in the network architecture?

   • AThe final fully connected layer (or classification head) must be replaced to match the new number of classes.
   • BThe pooling layers must be removed.
   • CThe input layer must be changed to accept the new image dimensions.
   • DThe convolutional layers must be replaced with new random filters.
   Reveal Solution Discussion Select an answer first
6. 25foundation · easy

   Why is the ReLU activation function commonly used in the convolutional layers of a CNN?

   • AIt bounds the output between -1 and 1, which helps stabilize training.
   • BIt introduces nonlinearity while being computationally efficient and helping mitigate the vanishing gradient problem.
   • CIt ensures that the gradient is always exactly 1 for all inputs.
   • DIt squashes the output to a probability distribution over classes.
   Reveal Solution Discussion Select an answer first