Pseudo-Labeling and Self-Training Quiz

Answers and explanations appear as you take the quiz.

1. How does a basic self-training loop use unlabeled examples?
   • A student labels only the original labeled set
   • A teacher predicts labels and selected predictions join true labels to train a student
   • Unlabeled examples replace the validation set
   • A teacher removes all uncertain samples from the world
2. Why apply a confidence threshold to pseudo-labels?
   • To force every unlabeled item into a class
   • To make all classes equally frequent automatically
   • To avoid training indiscriminately on uncertain predictions
   • To remove the need for validation
3. What is confirmation bias in pseudo-labeling?
   • A model ignores all unlabeled inputs
   • Labels leak from the test set automatically
   • The teacher and student have identical architectures
   • Early mistakes can be reinforced when the model trains on its own incorrect labels
4. In FixMatch, where does the pseudo-label come from?
   • A weakly augmented view, then it supervises a strongly augmented view if confidence is high
   • A strong view’s label supervises a weak view at any confidence
   • An unrelated image’s label is copied
   • The validation labels are added during training
5. What is the student-noise idea in Noisy Student?
   • Remove all augmentation from the student
   • Train a student with noise such as dropout or augmentation after pseudo-labeling
   • Add noise to the ground-truth labels only
   • Use a smaller student and no teacher
6. Why can class balancing be useful in pseudo-labeling?
   • Every threshold produces equal class counts
   • It makes labels independent of the teacher
   • High-confidence predictions may overrepresent some classes
   • It removes the need for real labels
7. What is the first model in a teacher-student setup usually trained on?
   • Only the test set
   • Pseudo-labels from a nonexistent teacher
   • Random class assignments for all inputs
   • The available labeled examples
8. What should be checked before using pseudo-labels to expand a medical image dataset?
   • Whether confident predictions and validation are appropriate for that task
   • Whether all images can be automatically treated as expert labels
   • Whether the model avoids all human review
   • Whether the test set can become training data