Cyclical Learning Rates Quiz

Answers and explanations appear as you take the quiz.

1. What core assumption do cyclical learning rates challenge?
   • That the learning rate should only ever decrease during training
   • That batch size must stay constant
   • That models need GPUs
   • That data must be shuffled
2. What did Smith report for cyclical learning rates on tested classification tasks?
   • They always lowered training loss to zero
   • They improved accuracy versus fixed rates, often in fewer iterations
   • They removed the need for validation data
   • They guaranteed escape from every sharp minimum
3. What does the 'LR range test' do?
   • Measures GPU memory usage
   • Counts the number of model parameters
   • Sweeps the learning rate upward while watching the loss to pick good bounds
   • Tests the model on out-of-distribution data
4. In the one-cycle policy, how does momentum typically behave relative to the learning rate?
   • Momentum stays fixed throughout
   • Momentum is set to zero the entire time
   • Momentum increases exactly in step with the rate
   • Momentum moves inversely — high when the rate is low and low when the rate is high
5. During a triangular cyclical learning-rate cycle, how does the rate move between its bounds?
   • It rises from the base rate to the maximum, then falls back toward the base
   • It only decreases from the first step
   • It stays fixed at the maximum
   • It jumps randomly at every minibatch
6. Who introduced cyclical learning rates and the LR range test?
   • Geoffrey Hinton
   • Leslie Smith
   • Yann LeCun
   • Andrej Karpathy
7. What sets the duration of a cyclical learning-rate cycle?
   • The number of model layers
   • The vocabulary size
   • A chosen number of training iterations for the up-and-down pattern
   • The number of optimizer parameters
8. What distinguishes the one-cycle policy from repeating several triangular cycles?
   • It keeps the learning rate constant throughout
   • It removes the optimizer’s momentum
   • It measures the learning rate only once
   • It uses one long learning-rate cycle for the training run