引導式測驗 · 中等 等級

視覺模型中的捷徑學習測驗

測試捷徑線索、源內與外部評估、受控背景變化、顯著性限制以及透過反例緩解。

問題 1 的 8

A wolf classifier relies on snow in training images. Why is that a shortcut for species recognition?

預覽每個問題8 問題

當您參加測驗時,就會出現答案和解釋。

  1. A wolf classifier relies on snow in training images. Why is that a shortcut for species recognition?
    • Snow can correlate with the label without defining the animal
    • Snow is a stable anatomical property of wolves
    • The camera directly measures species from background
    • Snow remains a species cue in every habitat
  2. Why can a random split from one camera campaign hide a source shortcut?
    • Random sampling guarantees a new capture source
    • Near-duplicate scenes are removed by randomization
    • The same source cannot appear in both partitions
    • Train and test can share the same camera-specific cues
  3. Which experiment best tests whether a model uses a suspect background?
    • Change the object and background together
    • Keep the target object while varying the background
    • Compare only two near-identical crops
    • Change model threshold without changing images
  4. A heat map highlights a corner watermark. What can the map establish alone?
    • The watermark definitely caused every prediction
    • The heat map measures the watermark’s causal effect
    • It suggests an area to investigate, but not causal reliance by itself
    • Removing the watermark must improve real-world accuracy
  5. Which data collection step may weaken a known backdrop shortcut?
    • Capture each target category across several backgrounds
    • Collect more duplicates of each class in one room
    • Keep one background for each label
    • Change category names without new images
  6. Why can more images from the same narrow source fail to solve shortcut learning?
    • More images automatically break the correlation
    • Training loss stops depending on nuisance features
    • The source cue disappears as data volume grows
    • They can reinforce the same nuisance-label correlation
  7. A hospital mark predicts a label in a research set. Which deployment check matters?
    • Treat the mark as a diagnostic clinical sign
    • Test on other hospitals where that mark does not track the outcome
    • Test only held-out images from the original site
    • Remove all labels for patients from the first site
  8. After removing a known shortcut, what conclusion is justified?
    • Every possible shortcut has been removed
    • The model necessarily understands the scene like a person
    • That one dependence may be reduced, while other shortcuts still require testing
    • Future source shifts no longer need monitoring