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。
- 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
- 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
- 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
- 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
- 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
- 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
- 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
- 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