준지도 학습 퀴즈의 일관성 정규화

퀴즈를 풀면 정답과 설명이 나타납니다.

1. What is the main consistency assumption?
   • Any change to an input must change its class
   • Unlabeled data should be assigned random labels
   • Small label-preserving perturbations should not change the model’s prediction
   • All examples should produce identical outputs
2. How is an unlabeled example used in a basic consistency loss?
   • Assign it a verified ground-truth label
   • Use it only to compute classification accuracy
   • Remove it after augmentation
   • Compare predictions from two perturbed views and penalize their difference
3. Which is an example of a label-preserving image perturbation when appropriate?
   • A modest crop or rotation that leaves the object class unchanged
   • Replacing the object with a different class
   • Removing all pixels containing the target
   • Changing the label at random
4. What does the low-density decision-boundary intuition suggest?
   • Place a boundary through every cluster center
   • Place decision boundaries away from dense groups of similar unlabeled examples
   • Ignore the unlabeled distribution
   • Force every cluster to have one class
5. What is a stop-gradient doing in a teacher-student consistency setup?
   • It removes the unsupervised loss
   • It freezes every model parameter permanently
   • It prevents the target branch from receiving the matching-loss gradient
   • It makes both views identical
6. How does FixMatch use confidence?
   • It uses every weak prediction as a hard label
   • It sets confidence to one for all classes
   • It drops the strongly augmented view
   • It keeps a pseudo-label from a weak view only when confidence passes a threshold
7. Why are unlabeled examples useful in this method?
   • They provide additional input structure without requiring a human label for each example
   • They guarantee the class labels are correct
   • They replace the labeled validation set
   • They eliminate the need for augmentation
8. Which evaluation would best test a consistency method’s intended behavior?
   • Check only training loss
   • Measure prediction stability across label-preserving views and task quality on held-out labeled data
   • Test on a single augmented view
   • Count unlabeled examples without measuring outputs