Kvíz spojený s průvodcem · Střední úroveň

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.

Související vodicí cestyGrad Cam And Saliency Maps
Otázka 1 z 8

What changes when Grad-CAM targets a different class score on the same image?

Náhled každé otázky8 otázky

Odpovědi a vysvětlení se zobrazují, jak budete v kvízu.

  1. 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
  2. Proč je mapa Grad-CAM obvykle hrubší než pixelová maska?
    • 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
  3. Nad vodoznakem se objevuje jasná oblast. Který další krok nejlépe prověří hypotézu o zkratce (shortcut)?
    • 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
  4. 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
  5. 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
    • Pouze velikost souboru obrázku
  6. Jaké důkazy podporují tvrzení, že lékařská teplotní mapa lokalizuje lézi?
    • 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
  7. Proč může být jediná intervence založená na oříznutí neprůkazná?
    • 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
  8. Které použití nejlépe odpovídá doporučené roli metody Grad-CAM v této příručce?
    • 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