CNN Visualizing CNN Feature Maps 4 — Questions and Answers
Question 1: Grad-CAM produces coarse localization maps. What is the primary reason for their low spatial resolution?
- It uses only the first convolutional layer
- It relies on the final convolutional layer which has small spatial dimensions due to pooling/striding (Correct answer)
- It averages across all layers equally
- It discards gradient information during backpropagation
Correct answer: It relies on the final convolutional layer which has small spatial dimensions due to pooling/striding
Grad-CAM uses the last convolutional layer's activations, which have been spatially compressed by pooling and striding, producing coarse heatmaps that are then upsampled.
Question 2: Guided Backpropagation improves standard backpropagation visualization by doing what?
- Backpropagating only through the final fully connected layer
- Zeroing out negative gradients at ReLU gates during the backward pass (Correct answer)
- Replacing all activations with their absolute values
- Using second-order gradients (Hessians) instead of first-order
Correct answer: Zeroing out negative gradients at ReLU gates during the backward pass
Guided backpropagation zeros gradients that are negative either in the upstream gradient OR in the forward activation, producing sharper, cleaner saliency maps.
Question 3: What is Guided Grad-CAM?
- Grad-CAM applied only to guided (adversarially trained) networks
- The element-wise product of Guided Backpropagation and Grad-CAM outputs (Correct answer)
- A technique that uses second derivatives to weight activation maps
- Grad-CAM restricted to the first convolutional layer
Correct answer: The element-wise product of Guided Backpropagation and Grad-CAM outputs
Guided Grad-CAM combines the class-discriminative localization of Grad-CAM with the fine-grained detail of Guided Backpropagation via element-wise multiplication.
Question 4: In feature map visualization, what does 'channel-wise mean activation' summarize?
- The average filter weight magnitude per channel
- The average activation value across all spatial positions in a feature map (Correct answer)
- The mean gradient flowing through a channel during backpropagation
- The average number of non-zero pixels in the input
Correct answer: The average activation value across all spatial positions in a feature map
Channel-wise mean activation collapses the spatial dimensions by averaging all positions in a feature map, giving a single scalar representing each channel's overall response.
Question 5: Which statement about LIME (Local Interpretable Model-agnostic Explanations) in the context of CNNs is correct?
- LIME directly visualizes convolutional feature maps
- LIME perturbs superpixels of the input and fits a local linear model to explain predictions (Correct answer)
- LIME requires access to the model's internal gradients
- LIME only works with fully connected networks
Correct answer: LIME perturbs superpixels of the input and fits a local linear model to explain predictions
LIME explains CNN predictions by masking image superpixels, observing prediction changes, and fitting a simple interpretable model to approximate local decision boundaries.
Question 6: When comparing feature maps of a correctly classified vs. misclassified image using the same CNN, what insight can this provide?
- It reveals which optimizer was used during training
- It shows which filters are responding differently, helping diagnose failure modes (Correct answer)
- It proves the model needs more training data
- It identifies overfitting in the training set exclusively
Correct answer: It shows which filters are responding differently, helping diagnose failure modes
Comparing activation patterns between correct and incorrect predictions highlights which intermediate representations differ, helping identify where the network's reasoning breaks down.
Question 7: What does the 'occlusion sensitivity' visualization method involve?
- Randomly dropping neurons during the forward pass
- Systematically masking patches of the input and measuring the change in model output confidence (Correct answer)
- Adding Gaussian noise to feature maps to test robustness
- Occluding gradient flow through specific filters during backprop
Correct answer: Systematically masking patches of the input and measuring the change in model output confidence
Occlusion sensitivity slides a gray patch over the input image and records how much the prediction score changes, revealing which regions are most critical for classification.
CNN Visualizing CNN Feature Maps 4
Grad-CAM produces coarse localization maps.
What is the primary reason for their low spatial resolution?