Guide-linked quiz · Medium Level
Grad-CAM and Visual Saliency Maps Quiz
Test Grad-CAM class targets, feature-map gradients, coarse localization, heat-map limits, controlled interventions, and expert validation.
Question 1 of 8
What changes when Grad-CAM targets a different class score on the same image?
Preview every question8 questions
Answers and explanations appear as you take the quiz.
- What changes when Grad-CAM targets a different class score on the same image?
- The original image acquires a new class label
- The model weights must be retrained first
- The resulting class-specific heat map can change
- The camera must capture another exposure
- Why is a Grad-CAM map usually coarser than a pixel mask?
- It stores only one heat value for the dataset
- It uses expert masks as its default input
- Upsampling creates new pixel-level evidence
- It comes from spatial feature maps that have lower resolution than the input image
- A bright region appears over a watermark. Which next step best tests a shortcut hypothesis?
- The map alone proves causal dependence
- Change or remove the watermark with controls and compare predictions
- Every highlighted pixel is a class probability
- A high color value certifies a correct prediction
- Why does a red Grad-CAM pixel not establish object-mask membership?
- The display is an upscaled contribution map, not an annotated segmentation
- Red pixels are annotated object ground truth
- Upscaling makes the feature grid pixel-accurate
- The map directly traces every model interaction
- Which information helps reproduce a Grad-CAM visualization?
- Only the screenshot display brightness
- Only the architecture’s public name
- Target output, selected layer and preprocessing choices
- Only the image file size
- What evidence supports a claim that a medical heat map localizes a lesion?
- A visually plausible overlay alone
- Comparison with independent expert region annotations
- A high class score without region labels
- Repeated maps from the same model alone
- Why can a single crop-based intervention be inconclusive?
- The crop may create an unnatural input shift as well as removing the cue
- Crops cannot change the input distribution
- A changed score proves only one feature mattered
- Cropping guarantees the intended label changes
- Which use best fits the guide’s recommended role for Grad-CAM?
- Use the map alone to certify clinical correctness
- Replace external validation with the overlay
- Treat every warm area as a verified lesion
- Generate audit hypotheses, then test them with other evidence
Keep testing yourself
More quizzes picked for your level.