Loss Functions Quiz

Answers and explanations appear as you take the quiz.

1. For a prediction task, what does its selected loss quantify against the training target?
   • A specified prediction-error objective
   • Only runtime speed
   • Only the number of network layers
   • Every application benefit automatically
2. Which loss function is most commonly used for classification tasks?
   • Mean squared error
   • Mean absolute error
   • Cross-entropy loss
   • Hinge distance only
3. Compared to mean squared error, why might you choose mean absolute error?
   • It squares errors to punish them more
   • It is more robust to outliers because it does not square errors
   • It cannot be used for regression
   • It ignores the true labels
4. Why does gradient-based training need a usable loss derivative or suitable substitute?
   • So a loss number can be printed
   • So all learning methods become identical
   • So the dataset no longer needs targets
   • So backpropagation and optimization can obtain an update signal
5. How does one-hot class cross-entropy change as probability assigned to the target class becomes very small?
   • Its loss must become zero
   • It rewards the prediction increasingly
   • Its loss becomes large
   • It becomes identical to accuracy
6. Why is accuracy usually not used directly as a training loss?
   • It is not differentiable, so it gives no useful gradient
   • It is too easy to compute
   • It only works for regression
   • It always equals the loss anyway
7. For illustrative regression residuals one and three, what are mean squared error and mean absolute error?
   • MSE is two and MAE is five
   • Both are four
   • MSE is ten and MAE is four
   • MSE is five and MAE is two
8. Using a base-two logarithm, what one-hot cross-entropy loss corresponds to target-class probability one-quarter?
   • One-quarter bit
   • Two bits
   • Four bits
   • Zero bits