Label Smoothing Quiz

Answers and explanations appear as you take the quiz.

1. What does label smoothing change about training targets?
   • It softens one-hot targets so the true class is below 1.0 and others are slightly above 0
   • It removes the loss function
   • It deletes incorrect classes
   • It doubles the learning rate
2. A common value for the smoothing parameter epsilon is:
   • 0.9
   • 0.1
   • 5.0
   • 0.0
3. What is one intended effect of replacing hard one-hot targets with smoothed targets?
   • Force every class to have equal probability
   • Guarantee the largest possible logit for the true class
   • Avoid requiring the labeled class to receive a target probability of exactly one
   • Remove the need for labeled training data
4. Which landmark architectures are prominent examples of label-smoothing use?
   • AlexNet and LeNet only
   • ResNet-50 and VGG only
   • AlphaGo and AlphaZero
   • Inception-v3 and the Transformer
5. What can label smoothing do to calibration?
   • It can improve calibration in some settings, but the effect is task- and model-dependent
   • It guarantees perfect calibration for every classifier
   • It cannot affect predicted probabilities
   • It always makes confidence equal accuracy
6. What is a documented downside of label smoothing?
   • It always reduces accuracy
   • It can erase fine between-class similarity, hurting knowledge distillation
   • It requires twice the training data
   • It only works on text
7. Why can label smoothing complicate knowledge distillation?
   • It converts every class label into an image
   • It forces the student to use a different tokenizer
   • It can remove information in logits about similarities among incorrect classes
   • It prevents the teacher from producing logits
8. What should a practitioner verify when comparing label-smoothing implementations?
   • Whether the model has a web browser
   • Whether training uses a validation filename
   • Whether logits are stored as text
   • Whether epsilon is distributed over all classes or only the non-target classes