Regularization Quiz

Answers and explanations appear as you take the quiz.

1. Why would a practitioner add regularization when a model fits training-specific noise?
   • To guarantee zero training error
   • To remove the need for validation data
   • To guarantee a faster GPU kernel
   • To discourage overfitting and evaluate whether unseen-data performance improves
2. Which regularization technique is most likely to drive some weights exactly to zero, effectively selecting features?
   • L2 regularization (weight decay)
   • L1 regularization
   • Batch normalization
   • Increasing the learning rate
3. How does dropout regularize a neural network during training?
   • It deletes layers permanently from the network
   • It adds the squared weights to the loss function
   • It randomly switches off neurons each step so they cannot co-adapt
   • It stops training when validation loss rises
4. In the regularized loss 'error + lambda × penalty', what does increasing lambda do?
   • Strengthens the penalty, pushing the model toward simpler solutions
   • Weakens the penalty and allows larger weights
   • Removes the error term entirely
   • Has no effect on the weights
5. Why can early stopping be considered a form of regularization?
   • Because it adds a weight penalty to the loss
   • Because it halts training before the model starts fitting noise in the data
   • Because it increases model capacity over time
   • Because it randomly removes neurons
6. What can happen when regularization is strengthened beyond what a task needs?
   • Both training and test accuracy must improve
   • The penalty becomes unrelated to the objective
   • Validation becomes unnecessary
   • The model can underfit useful structure even if variance decreases
7. For a regularization example with weights minus two and one, what are the unscaled L1 and sum-of-squares L2 penalties?
   • L1 is three and L2 is five
   • L1 is minus one and L2 is three
   • L1 is five and L2 is three
   • Both penalties are minus two
8. A regularized objective has prediction error two, lambda one-half, and penalty six. What is its total value?
   • Eight
   • Three
   • Five
   • One