Sharpness-Aware Minimization Quiz

Answers and explanations appear as you take the quiz.

1. What kind of minimum does SAM try to find?
   • A flat minimum where loss stays low across a neighborhood of weights
   • The sharpest possible minimum
   • The minimum with the most parameters
   • Any minimum reachable in one step
2. How many forward-backward passes does a standard SAM step require?
   • Half a pass
   • Roughly two
   • Exactly one
   • Ten or more
3. What does the radius hyperparameter rho control in SAM?
   • The learning rate decay speed
   • The number of training epochs
   • The size of the neighborhood around the weights that is protected against loss spikes
   • The dropout probability
4. What is the first of SAM's two steps at each iteration?
   • Shrink the model
   • Average the weights of all past steps
   • Reset the optimizer state
   • Take an ascent step of size rho toward the worst-case nearby point
5. Why does SAM often improve robustness to label noise?
   • Because flat minima are less prone to memorizing corrupted labels
   • Because it deletes noisy labels automatically
   • Because it uses a larger vocabulary
   • Because it triples the batch size
6. Which of these is a variant designed to reduce SAM's compute cost?
   • ReLU
   • ESAM (Efficient SAM)
   • BatchNorm
   • AdaGrad
7. In SAM’s second pass, which gradient is used to update the original model weights?
   • The gradient of a random label
   • The first pass gradient without evaluating the perturbed point
   • The gradient evaluated at the perturbed nearby weights
   • A gradient from a different model checkpoint
8. If SAM’s neighborhood radius rho is set to zero, what happens to its perturbation step?
   • The model parameters are all set to zero
   • The loss becomes independent of the weights
   • The optimizer takes an infinite step
   • There is no weight perturbation, so the neighborhood reduces to the current point