Semi-Supervised Learning Quiz

Answers and explanations appear as you take the quiz.

1. Semi-supervised learning is designed for situations where:
   • Every example is fully labeled
   • You have no data at all
   • Labels are infinitely cheap to obtain
   • You have a little labeled data and a lot of unlabeled data
2. How does self-training use selected predictions on unlabeled examples?
   • It treats selected model guesses as pseudo-labels for further fitting
   • It converts every guess into a verified human annotation
   • It permanently removes the original labeled examples by definition
   • It chooses labels only from alphabetical order
3. Under consistency regularization, what agreement is encouraged across chosen class-preserving augmentations?
   • A requirement that every arbitrary edit preserve the class
   • Opposite predictions for every view of an example
   • Agreement in the model's predictions
   • Removal of every unlabeled input from the loss
4. What does a cluster assumption propose for grouped observations in a semi-supervised method?
   • Every possible cluster must contain every class equally
   • They are likely to share a label under the assumed data structure
   • All semi-supervised methods guarantee identical clusters
   • Labels are determined by filename order
5. How does FixMatch combine the two main ideas?
   • It pseudo-labels a weakly augmented input, then trains the strongly augmented version to match
   • It uses only labeled data and ignores everything else
   • It deletes augmentations entirely
   • It assigns labels purely at random
6. How can confirmation bias arise during pseudo-label training?
   • Only low-confidence labels can ever be wrong
   • Every selected prediction becomes a verified fact
   • Using any verified label prevents the model from learning
   • Incorrect model-generated labels are accepted and then reinforced
7. Which structure distinguishes graph-based label propagation from repeatedly selecting a classifier's confident guesses?
   • Only a threshold on independently predicted class probabilities
   • A similarity graph over labeled and unlabeled observations
   • A list of random output labels with no input similarities
   • A requirement to discard every relationship between observations
8. In the stated scikit-learn semi-supervised interface, which integer marks an unlabeled target entry?
   • Zero, necessarily the first actual class label
   • One, necessarily the second actual class label
   • Minus one
   • Fifty, interpreted as a percentage-confidence threshold